MobileMAN Presentation

Document technical information

Format pdf
Size 8.2 MB
First found May 22, 2018

Document content analysis

Category Also themed
not defined
no text concepts found


Ike Skelton
Ike Skelton

wikipedia, lookup




Mobile Metropolitan Ad hoc Networks
MobileMAN Presentation
Deliverable D15
Contractual Report Preparation Date: July 2005 (*)
Actual Date of Delivery: 29 July 2005
Estimated Person Months: 7.5
Number of pages: 149
Contributing Partners: Consiglio Nazionale delle Ricerche (Italy),
University of Cambridge (UK), Institut Eurecom (France), Helsinki
University of Technology (Finland), NETikos (Italy), Scuola
Universitaria Professionale della Svizzera Italiana (Switzerland)
Authors: Marco Conti, Franca Delmastro, Enrico Gregori, Antonio
Pinizzotto (CNR), Jon Crowcroft, Andrea Passarella (Cambridge),
Claudio Lavecchia, Pietro Michiardi, Refik Molva (Eurecom), Jose
Costa Requena, Mohammand Ayyash (HUT), Piergiorgio Cremonese,
Veronica Vanni (Netikos) Ralph Bernasconi, Claudia Brazzola, Ivan
Defilippis, Jennifer Duyne, Silvia Giordano, Alessandro Puiatti,
(*) new schedule accepted by the Project Officer
Abstract: The aim of this deliverable is to report on the international workshop we organized to
disseminate the MobileMAN activities inside the scientific and industrial community. To this
end, in the framework of the IEEE International Conference on Pervasive Services 2005
(ICPS'05) we organized the First IEEE ICPS Workshop on Multi-hop Ad hoc Networks: from
theory to reality, REALMAN 2005 (
Project funded by the European Community
under the “Information Society
Technologies” Programme (1998-2002)
July 2005
The REALMAN workshop idea steams from the lessons we learnt in the framework
of the MobileMAN project. The results obtained in the MobileMAN project and the
emerging world-wide literature indicated that to consolidate the ad hoc networking
field we need to complement theoretical research activities with the realization and
testing of realistic small/medium scale testbed. Indeed, after almost a decade of
research into ad hoc networking, MANET technology has not yet affected our way of
using wireless networks and there are no clear results showing how well MANETs
work in reality as the research in this area has been based on theoretical analyses and
simulation results only. Simulation models often introduce simplifications and
assumptions that mask (in simulation experiments) important characteristics of the
real protocols behavior. To make mobile multi-hop ad hoc networks (MANETs) a
reality, simulation modeling and theoretical analyses have to be complemented by real
experiences (e.g., measurements on real prototypes) which provide both a direct
evaluation of ad hoc networks and, at the same time, precious information for a
realistic tuning of simulation models. In addition, the availability of prototypes will
also make possible to start creating communities of MANET users that, by
experimenting with this technology, will provide feedbacks on its usefulness and
stimulate the development of applications tailored for the ad hoc environment. This is
fundamental to reduce the gap between what end users might find useful, and what
research is currently addressing, making the cost of using ad hoc networking lower
than the potential benefit.
The need for more experimental activities has stimulated the emerging of a
new community of researchers combining theoretical research on ad hoc networking
with experiences/measurements obtained by implementing ad hoc network
prototypes. The aim of REALMAN was to bring together this community and to
disseminate the MobileMAN results to the scientific community. The workshop
constituted a unique forum for presenting and discussing experiences/results from real
ad hoc networks test-beds and prototypes.
The answer to the open call for papers was very encouraging. We received 39
submissions out of which we selected 13 papers for presentation in the workshop
sessions. In addition, in response to a separate call we received several interesting
Deliverable D15
2 of 5
July 2005
demo proposals. The final program also included 8 demos (5 from MobileMAN
partners) and 4 posters. 36 worldwide researchers participated to the workshop. These
numbers show that REALMAN is already a reference forum for the growing
community of researchers working in this field.
The workshop also included a panel “How to make MANETs scale and
provide coverage for real” to discuss/identify research directions for making mobile
ad hoc networks a reality. The panel, organized and chaired by Jon Crowcroft, had a
set of panelists -- Matthias Grossglauser (EPFL), Edward Knightly (RICE University),
Martin Mauve (Duesseldorf University), Joerg Ott (Helsinki University of
Technology), Christian Tschudin (Basel University) -- that addressed the discussion
from two different perspectives. Specifically, Matthias Grossglauser started by
observing that applications are hard to predict (and hence scale is easy to
underestimate); he pointed out the value of model-based research, especially for
architectural decisions. On the opposite, the other speakers emphasized the key role of
applications and real implementations. Martin Mauve took the opposing view to
Matthias Grossglauser: application led research is better! Martin gave many examples
from a large vehicle/car-to-car network they built, telling how the wireless network
design worked much better when the architecture was application centered.
Edward Knightly and Joerg Ott discussed two directions for the evolution of
ad hoc networks: mesh networks and opportunistic networking. Edward Knightly
discussed research challenges in implementing a real and scalable mesh network by
exploiting his experience with the RICE TAP project; Joerg Ott talked about
opportunistic networking and pointed out very nice real examples to show that
opportunistic networking (i.e., delay tolerant networks, DTN) can immediately
provide useful applications for ad hoc networking. In Joerg view DTN precedes
Christian Tschudin summarized the discussion pointing out the need for a
pragmatic approach in the mobile ad hoc research: use an application driven
approach, and integrate/consolidate existing algorithms and software. Specifically, he
stated: While in theory it might be feasible to have 100 hop paths, I don't think that
this will happen in real MANETs. The limit lies somewhere below and is determined
more by pragmatic property bundles like routing stability or TCP fairness rather than
single functions (finding a route) or optimizations (battery life time). What can theory,
Deliverable D15
3 of 5
July 2005
what can systems work contribute to bracket this ad hoc horizon? Agreeing on the
"real objectives" would be a first step.
The slides of the panel presentations and other workshop material can be
found at:
The workshop proceedings are in the attached Appendix.
Deliverable D15
4 of 5
July 2005
This Appendix contains the referred proceedings of the First IEEE ICPS Workshop on
Multi-hop Ad hoc Networks: from theory to reality, REALMAN 2005
Deliverable D15
5 of 5
Table of Contents
Organizing Committee
Session I: MANET Implementations
Experiments with an enhanced MAC architecture for multi-hop wireless networks
Ralf Bernasconi, Raffaele Bruno, Ivan Defilippis, Silvia Giordano, and Alessandro Puiatti
Experiments of Ana4: An Implementation of a 2.5 Framework for Deploying Real Multi-hop Ad
hoc and Mesh Networks
Nicolas Boulicault, Guillaume Chelius, and Eric Fleury
GNU/Linux Implementation of a Position-based Routing Protocol
Marc Heissenbüttel, Torsten Braun, Tobias Roth, and Thomas Bernoulli
Implementation Strategies for a Secure and Efficient Multi-hop MANET Platform
Minmin Tu, Jingyu Zhou, and Guozhi Xu
A linux based Bluetooth scatternet formation kit: from design to performance results
Francesca Cuomo and Andrea Pugini
Session II: MANET Experimentations
An Experimental Study of P2P Group-Communication Applications in Real-World MANETs
Franca Delmastro and Andrea Passarella
A Comparative Study of Cooperative Algorithms for Wireless Ad Hoc Networks
Alan Lim, Vikram Srinivasan, and Chen-Khong Tham
A Path Density Protocol for MANETs
Evgeny Osipov and Christian Tschudin
Interactions between TCP, UDP and Routing Protocols in Wireless Multi-hop Ad hoc Networks
Christian Rohner, Erik Nordström, Per Gunningberg, and Christian Tschudin
Hop of No Return: Practical Limitations of Wireless Multi-Hop Networking
Marina Petrova, Lili Wu, Matthias Wellens, and Petri Mahonen
Session III: Deploying MANET Test-beds
Thoughts on Mobile Ad-hoc Network Testbeds
Wolfgang Kieβ, Stephan Zalewski, Andreas Tarp, and Martin Mauve
Experiences Deploying an Ad-hoc Network in an Urban Environment
Peter Barron, Stefan Weber, Siobhán Clarke, and Vinny Cahill
MeshDV: A Distance Vector mobility-tolerant routing protocol for Wireless Mesh Networks
Luigi Iannone and Serge Fdida
Posters and Demos Session
Social networks, novel communication applications and needs in mobile contexts
Claudia Brazzola
On the Dimensionality of Wireless Connectivity Traces
George Roussos
Cross-Layer Support for Group-Communication Applications in MANETs
Marco Conti, Franca Delmastro, Jon Crowcroft, and Andrea Passarella
Real Life Experience of Cooperation Enforcement Based on Reputation (CORE) for MANETs
Claudio Lavecchia, Pietro Michiardi, and Refik Molva
VoIP Testbed in Ad Hoc Networks
Jose Costa-Requena, Mohammand Ayyash, Jarrod Creado, Jarkko Hakkinen, Raimo Kantola, and
Nicklas Beijar
Experimenting a Layer 2-based Approach to Internet Connectivity for Ad Hoc Networks
Raffaele Bruno, Marco Conti, Enrico Gregori, Antonio Pinizzotto, and Emilio Ancillotti
Demo of Ana4: an Hybrid Local Area Ad hoc Network Architecture
Nicolas Boulicault, Guillaume Chelius, and Eric Fleury
Haggle Architecture and Demo of its Real World Implementations
Pan Hui, Jon Crowcroft, James Scott, Christophe Diot, Augustin Chaintreau, and Richard Gass
Demo of residual bandwidth estimation in an 802.11 ad hoc network
Martin Nielsen
Organizing Committee
General Chair:
Jon Crowcroft, University of Cambridge, UK
Program Chair:
Marco Conti, National Research Council, IIT Institute, Italy
Program Vice-Chair:
Andrea Passarella, University of Cambridge, UK
Technical Program Committee: G. Anastasi, Pisa University, Italy
A.T. Campbell, Columbia University, USA
S.R. Das, SUNY at Stony Brook, USA
A. Ephremides, University of Maryland, USA
S. Giordano, SUPSI, CH
E. Gregori, IIT-CNR, Italy
T. Henderson, Dartmouth College, USA
R. Kantola, Helsinki University of Technology, Finland
H. Karl, Paderborn University, Germany
E. Knightly, Rice University, USA
M. Mauve, Dusseldorf University, Germany
P. Michiardi, EURECOM, France
R. Molva, EURECOM, France
S. Olariu, Old Dominion University, USA
G. Polyzos, AUEB, Greece
D. Remondo, Catalonia University of Technology, Spain
C. Rohner, Uppsala University, Sweden
J.H. Schiller, Freie University Berlin, Germany
J. Scott, Intel Research, UK
F. Sestini, European Commission
V.A. Siris, FORTH-ICS and Crete University, Greece
I. Stavrakakis, Athens University, Greece
I. Stojmenovic, Ottawa University, Canada
C. Tschudin, Basel University, Switzerland
N. Vaidya, University of Illinois at Urbana-Champaign,USA
J. Wu, Florida Atlantic University, USA
MANET Implementations
Experiments with an enhanced MAC architecture for multi-hop wireless
R. Bernasconi 1 , R. Bruno 2 , I. Defilippis 1 , S. Giordano 1,2 , A. Puiatti 1
1: University of Applied Sciences, (SUPSI), Department of Technology and Innovation (DTI),
6928 Manno, Switzerland
{ralph.bernasconi, ivan.defilippis, silvia.giordano, puiatti}
2: National Research Council, (CNR), IIT Institute, 56100 Pisa, Italy
[email protected]
In this paper we describe the architecture of a
wireless network interface card that we have used for
implementing an enhanced IEEE 802.11 MAC protocol for multi-hop wireless networks, which adopts a
backoff mechanism, called Asymptotically Optimal
Backoff (AOB) [5], specifically designed for high contention conditions. We have performed traffic experiments with this enhanced card to compare it against
the traditional IEEE 802.11 MAC protocol. The results of our experiments, mainly devoted to conflict
and interferences situations, show that our enhanced
version of the IEEE 802.11 MAC protocol is more
robust to both high contention and interferences situations than the standard one, resulting in a significant
improvement of per-station throughput performance1 .
1. Introduction
Instantly deployable, multi-hop wireless networks
can be both fully self-organized (ad hoc networks) and
connected to the wired backbone via multiple wireless
hops (multi-hop wireless LANs). Potential applications of this type of networks include rescue, citizen
and military networks, where users self-organize to
communicate or to access the Internet; and multi-hop
Hot Spots, i.e., wireless networks in public areas
where traditional WLAN technology is augmented
with ad hoc wireless communications.
Today, the de facto technology for building multihop ad hoc networks is the IEEE 802.11 standard [1].
However, it is extensively recognized that the 802.11
1 This
work was partially funded by the Information Society Technologies programme of the European Commission, Future and
Emerging Technologies under the IST-2001-38113 MOBILEMAN project.
standard backoff algorithm significantly degrades the
channel utilization in conditions of high contention,
because this policy has to pay the cost of collisions to
increase the backoff time when the network is congested [2-6]. Several feedback-based mechanisms have
been proposed to maximize the IEEE 802.11 MAC
protocol efficiency by guaranteeing an optimal time
spreading of the users’ access [2], [4-6]. Specifically,
in [5] the Asymptotically Optimal Backoff (AOB)
mechanism was proposed, which dynamically tunes
the backoff parameters such as to avoid that the network contention level exceeds its optimal value. The
AOB scheme estimates the network contention level
through the measure of the utilization rate of slots.
In a previous paper [14], we proposed the architecture of an enhanced IEEE 802.11 wireless network
interface card, which is capable of seamlessly working
in a multi-hop wireless network, but, at the same time,
it is still fully compatible with traditional IEEE
802.11 implementations. This Medium Access platform has been designed to be a suitable architecture for
implementing and testing: 1) backoff algorithms more
adequate to multi-hop operations; 2) dynamic channel
switching schemes to exploit channel quality diversity;
3) efficient packet forwarding schemes inside the MAC
layer; and 4) cross-layering optimizations through the
exploitations of topology information provided by the
routing layer.
In this paper, we present our activity concerning the
point 1) above. Specifically, we describe our first card
implementation, with the introduction of the AOB
scheme as described in [5]. We decided to implement
the AOB algorithm, because it relies only on topology-blind estimates of the network status based on the
standard physical carrier sensing activity. Hence, it
appears as a suitable solution for multi-hop configurations. The experimental results obtained comparing our
enhanced MAC card with traditional IEEE 802.11
wireless cards, show the significant per-station
throughput improvement ensured by the enhanced
MAC protocol. Furthermore, they open promising
directions to investigate additional enhancements, as
discussed in Section 4.
The rest of this paper is organized as follows: in
Section 2 we outline our implementation and the main
hardware and firmware design choices. In Section 3 we
describe the measurement environment and we present
the results of our real experiments, discussing the most
relevant points. Section 4 concludes this paper with
some further discussion and detailed description of the
ongoing and future work.
2. Overview of the Wireless Network Interface (WNI)
The Wireless Network Interface (WNI) is a multicomponent system that has been designed to implement and test enhanced MAC protocols. It is portable,
and therefore suitable for mobile use, provided that the
required power can be supplied.
setup, very similar to the one used in the experiments
presented in this paper.
The system may be used in lab environment (using
the JTAG interface to build the MAC firmware and to
download it on the DSP on-board FLASH memory),
as for instance during synthetic traffic tests. It may
also be used in a real environment, using the high
speed IEEE1394a (FireWire) bus, which allows the full
speed connection with a host PC running the user applications (not shown in Figure 2).
The RF and BB components have been kept compliant to the 802.11 standard, such as to allow mixed
environment experiments, where enhanced systems cooperate with standard, off-the-shelf IEEE 802.11 wireless cards. For the tests, 4 enhanced systems have been
2.1. Hardware overview
The selection of the hardware was a critical phase in
the realization of the WNI, and it required an extensive
analysis of components currently available on the market. We had to face different problems in order to be
able to find components that could be combined, and
to have the required instruments for realizing the
802.11 MAC, maintaining the flexibility needed to
support various experiments and allow future extensions.
Micro -line C 6713Compact
DSP / FPGA / IEEE 1394 board
DT 20 Modem
Direct Sequence
Spread Spectrum
Bypassed device
Figure 1: The enhanced IEEE 802.11 Wireless
Network Interface (WNI).
The system consists of Radio Frequency up/down
converters (RF), the Baseband modulator/demodulator
(BB), and the MAC firmware on a compact DSP microcontroller.
Figure 2: Typical laboratory setup.
Figure 1 shows the WNI in a laboratory environment, and Figure 2 illustrates a typical laboratory
Figure 3: Overview of the enhanced 802.11 Wireless
Network Interface (PHY).
The final selection was for the Elektrobit DT20
modem and the Orsys Micro-line C6713Compact DSP
board. The Elektrobit DT20 modem performs the BB
and the RF processing in compliance with the 802.11
PHY layer using an Intersil Prism-I WLAN chipset,
while the Orsys Micro-line C6713Compact DSP board
is used for the MAC implementation. RF modem and
DSP board are connected through a small custom-made
adapter board, which performs a voltage level conversion to adapt the logic signals voltage levels (3.3V for
the DSP board and 5V for the modem).
The DSP board integrates a large FPGA (Xilinx
XC2V250) and a Texas Instruments TMX320C6713
DSP. The FPGA has been used for interface adaptation
between DSP and Intersil components, and could be
used in the future in order to accelerate other tasks
(e.g., address filtering, cryptography, etc.). The 802.11
MAC firmware is developed in standard C, but specifically customized for the C6713 DSP. A more detailed overview of the interfaces between the logic
components constituting the MAC implementation is
depicted in the block diagram shown in Figure 4.
Texas Instruments
TMX 320 C 6713
XC 2 V 250
HFA 3824 A
RX /TX interface
HFA 3824 A/
HFA 3524
Control Port
Control Port
64 - bit
HFA 3824 A
Direct Sequence
Spread Spectrum
HFA 3524
Figure 4: Logic blocks diagram of the MAC implementation.
2.2. Firmware overview
The firmware running on the C6713 DSP is basically a simple monitor loop with few interrupt routines. However, an operating system was not needed
for the implementation of the standard 802.11 Frame
Exchange Sequence and relative tasks (fragmentation,
de-fragmentation, fragment cache control, etc.). Nevertheless, a Real-Time operating system could be easily
installed to accommodate more sophisticated procedures. Thus, in the software we have developed, only
few components are specific to the C6713Compact
board; among these: timing considerations, available
DSP resources, configuration and control required for
the specific implementation. On the actual system
(C6713Compact board), the firmware occupies about
125 Kbytes and can reside completely in the DSP internal RAM, at run time. It is worth pointing out that
the source coded we have developed is highly portable
because it is not tailored to a specific OS.
Since the original DSP implemented only a basic
monitor loop, we had to develop our MAC protocol
implementation starting from scratch. We designed our
software maintaining the maximum possible abstraction, in order to minimize the software re-design in
case of change of the development platform, and to
guarantee sufficient flexibility for future extensions. It
is worth remarking that there is still a large margin left
for the DSP, because the C6713 DSP could easily
accommodate more demanding standards such as the
11 Mbps IEEE802.11b.
The firmware we have developed is subdivided into
the following components:
MAC Firmware. This is the hard real-time software, which allows packets (fragments) to be
physically transmitted to and received from the
RF interface.
Host Interface Firmware. This software component managed the communication flows between
the host CPU and the DSP. This communications are carried out through the on board
IEEE1394a port.
Packet Data Management. To manage the data
flow between the communication channel and the
host, specialized data structures are needed. Given
the hard time constraints of the MAC firmware
these data structures must be optimized in terms
of numbers of memory access, numbers of data
swaps and storage.
The code implementing the 802.11 MAC software
is composed of 13 modules (source & header files).
Their functionalities and described in [8], but are omitted here for space constraints. Finally, the firmware is
linked with the c6x_boardlib, which is a third party
library implemented by Orsys for the C6713Compact
2.3 Enhanced backoff
In this section we outline the operations of the
AOB mechanism [5], which has been used for implementing the enhanced IEEE 802.11 MAC protocol in
our WNI. The AOB scheme was proposed to improve
the efficiency of the IEEE 802.11 MAC protocol, by
extending the standard binary exponential backoff to
guarantee that the wireless stations adopt the optimal
backoff interval corresponding to the current contention
level in the network. The utilization rate of the slots
(also denoted as slot utilization) is used as an estimate
of the current network contention level. The slot utilization can be computed as the ratio between the number of slots in the backoff interval in which one or
more stations start a transmission attempt, i.e., busy
slots, and the total number of slots available for
transmission in the backoff interval, i.e., the sum of
idle slots and busy slots. The slot utilization definition wasn’t originally introduced by the AOB mechanism. In [4], the DCC mechanism was proposed that
exploits the slot utilization ( SU ) to decide in a probabilistic way when performing a transmission attempt
or further deferring the access to the channel. Specifically, the DCC mechanism computes a Probability of
Transmission P _T according to the following formula:
P _T = 1 SU N _ A ,
where N _ A is the number of unsuccessful transmission attempts already performed by the station for the
transmission of the current frame. When the standard
802.11 MAC protocol assigns a transmission opportunity to a station (i.e., that station has backoff timer
equal to zero and perceive the channel as idle), the station will perform a real transmission with probability
P _T ; otherwise (i.e., with probability 1 P _T ) the
station deems the transmission opportunity as a virtual
collision, and the frame transmission is rescheduled as
in the case of a real collision, i.e., after selecting a new
backoff interval. By using the P _T defined in Equation (1), the DCC mechanism is capable of limiting
the slot utilization value.
Numerical results presented in [4] indicate that the
DCC mechanism is effective in reducing the contention level, and this is beneficial to increase the channel
utilization in 802.11 WLANs. However, the DCC
solution operates in a heuristic way, using larger contention windows than the standard protocol when the
contention increases. The AOB mechanism steps forward because it exploits the analytical characterization
of the IEEE 802.11 MAC protocol carried out in [11]
and [5] to identify the optimal slot-utilization level the
network should obtain to guarantee the maximum
channel utilization. Specifically, in [5] it was derived
the Asymptotic Contention Limit ( ACL ) function,
which approximates the optimal slot utilization with
an accuracy that increases as the network contention,
i.e., the number of active stations and/or message
length, increases. By exploiting the knowledge of the
ACL value, the AOB mechanism generalizes the expression of the P _ T parameter introduced in the DCC
scheme as follows:
S _U N _ A
P _T = 1 min1,
ACL (2)
The P _T expression defined in Equation (2) implies
that the probability of performing a transmission attempt tends to zero as the slot utilization approaches
the optimal contention level. The AOB scheme guarantees that the system operates asymptotically close to
the ACL value, i.e., that the channel utilization is
maximal in networks with a large number of stations.
It is worth pointing out that the analytical characterization of the ACL values presented in [4], demonstrated that the optimal value is almost independent of
number of active stations, but depends heavily on the
average message length. Hence, the AOB mechanism
guarantees to obtain the maximum channel utilization
without any knowledge of the number of active stations, overcoming the limitations of previous solu-
tions that adapt the backoff value to the network contention level [2], [6].
The enhanced IEEE 802.11 MAC protocol we have
implemented is a slight variation of the original AOB
scheme, although it is totally equivalent as far as the
protocol behaviour and performance. We decided to do
not estimate the aggregate slot utilization, as done in
[4], [5], but we split it into two contributions: the
internal slot utilization ( SU int ) and the external slot
utilization ( SU ext ), such as to differentiate between the
contribution to the channel occupation due to the
node’s transmissions and to its neighbors’ transmissions. This is motivated by the need to keep our implementation as much flexible as possible, in order to
allow future modifications. In fact, several extensions
of the AOB scheme are currently under investigation
[12], and the implementation of some of these proposals is an ongoing activity. Another variation with respect to the original AOB is the time interval over
which we compute the slot utilization. In fact, the
original AOB computes the slot utilization after each
backoff interval, while in our implementation we used
a constant observation period T equal to 100ms. This
choice is motivated by the need to avoid frequent slot
utilization computations, which could interfere with
the time constraints of the atomic MAC operations
(e.g., RTS/CTS exchange). Thus, considering a fixed
observation period T , each station computes the internal slot utilization, SU int , and the external slot utilization, SU ext , as:
SU int =
SU ext =
N tx
T idle
+ N rx + N tx
t slot
N rx
T idle
+ N rx + N tx
t slot
where N tx is the number of transmissions carried
out by the station during the T period; N rx is the
number of separate channel occupations (either successful transmission or collisions) observed during the T
period; and T idle is time the channel is idle during the
T period (including the idle periods during which the
DIFS and EIFS timers are active). It is evident that
the SU used in Equation (1) and Equation (2) can be
computed as the sum of SU int and SU ext . Thus, our
implementation and the original AOB are equivalent.
3. Measurement and Results
In order to validate our enhanced architecture we
carried out comparative tests of the performance
achieved by the legacy IEEE 802.11 backoff mecha-
nism and the enhanced one. In both sets of experiments we used our WNI implementation. All the tests
were performed in a laboratory environment, considering ad hoc networks in single-hop configurations.
Nodes were communicating in ad hoc mode and the
traffic was artificially generated. In our scenarios we
used a maximum of 4 stations, due to hardware limitations. However, this is not seen as a problem, because
we were already able to demonstrate the performance of
our solution and the coherence with previously performed simulations. In Section 3 we further elaborate
on this point, providing simulation results to support
our claim.
As discussed in Section 2, the average backoff
value that maximizes the channel utilization is almost
independent of the network configuration (number of
competing stations), but depends only on the average
packet sizes [11]. Therefore, the ACL value for the
frames size used in our experiments can be precomputed and loaded in the MAC firmware. The implementation in the FPGA of the algorithm defining
the ACL value in order to compute it at run-time, is an
ongoing activity.
periments was carried out to verify this performance
decrease in network configurations where two stations
are performing either a unidirectional or bidirectional
communication, as illustrated in Figure 6.
20 cm
MAC Tester via RS 232
The results we obtained in this point-to-point configuration are reported in Table I. In the table, N tx
denotes the average number of transmission (either
successes of collisions) performed during a period T ,
while RC denotes the average number of collisions
suffered during a period T . Hence, the average
throughout TP , expressed in byte/s, can be computed
N tx RC
From the listed numerical results, we can observe that
the throughput decrease in the case of two competing
stations is lower than 3%.
Table I: Point-to-point scenario results.
Unidirectional Flow
Bidirectional Flow
Standard Enhanced
30 cm
60 cm
In the second set of experiments we considered a
network configuration with 3 stations, as depicted in
Figure 7.
20 cm
Figure 6: Unidirectional and Bidirectional communications with two stations.
TP =
We used different scenarios (2, 3 and 4 stations), in
order to study at the same time, the performance of our
implementation and the correspondence with the previous simulation results. The stations were identically
programmed to continuously send 500-bytes long
MSDUs (MSDU denotes the frame payload). The consecutive MSDU transmissions were separated by at
least one backoff interval and we did not use the
RTS/CTS handshake, or the fragmentation. The
minimum contention window was set to 8 t slot (160
µsec), and all values were computed in stationary conditions. The nodes topology is illustrated in Figure 5.
All the experimental results we show henceforth were
obtained by computing the average over ten replications of the same test.
MAC Tester via RS 232
3.1. MAC Performance tests
Figure 5: nodes topology used in the measurements.
As already demonstrated in [5] and [12] the AOB
mechanism introduces a minimum overhead that could
negatively affect the performance of the communications between two stations. Thus, our first set of ex-
MAC Tester via RS 232
Figure 7: 3 stations scenario.
The experimental results we obtained in the 3 stations configuration are reported in Table II. We can
note that with three competing stations, the throughput
improvement is about 5.7%. This is explained by observing that the number of collisions occurred during a
T period is four times less for the enhanced MAC
than for the standard one.
Table IV: Summary of experimental results: total
throughput (Kbyte/s).
2 STAs (bidir.)
3 STAs
4 STAs
Table II: 3 stations scenario results.
Standard MAC
Enhanced MAC
54485 (+5.7%)
Finally, the last set of experiments was carried out
in the 4 stations scenario depicted in Figure 8.
MAC Tester via RS 232
3.2. Relations with previous simulation
We carried out experiments with up to four stations
due to hardware limitations. However, we argue that
the positive trend observed in our tests will be confirmed also for large numbers. To substantiate this
statement, in this subsection we show numerical results obtained through discrete-event simulations, considering the same parameter setting adopted during the
real tests. In particular, Figure 9 shows the channel
utilization of the IEEE 802:11 MAC protocol (rate
2Mbps) with and without the AOB mechanism, versus
the number of wireless stations in the network for an
ideal wireless channel, i.e., not affected by noise. In
the same figure we also show the maximum throughput achieved when the stations adopts the optimal
backoff interval (computed according to [6]). The
shown results refer to a payload size of 500 bytes, and
were obtained using a minimum contention window
equal to 8 time slots, as in our real experiments.
Figure 8: 4 stations scenario.
Table III: 4 stations scenario results.
Standard MAC
Enhanced MAC
42300 (+9.4%)
To summarize, in Table IV we report the aggregate
throughput measured in the different network configurations we have tested. The numerical results clearly
demonstrate that the enhanced MAC protocol may
significantly improve the per-station throughput as the
number of stations increases.
Throughput (byte/s)
The experimental results obtained in the 4 stations
configurations are reported in Table III. These results
confirm the positive trend shown in the previous experiments, since the throughput increase in the case of
four stations is 9.4%.
Protocol Capacity
STD 802.11
Number of Wireless Stations
Figure 9: Throughput of the IEEE 802.11 protocol
with and without the AOB mechanism versus optimal value.
From the figure, it is straightforward to note that
the throughput measured in the real tests is always
lower than the one obtained through simulations. The
worse throughput performance is due to the negative
impact of the radio interferences that are present in a
real environment, and this is true for both the standard
802.11 MAC protocol and the enhanced one. However,
the comparison between the experimental results reported in Table IV, and the simulation results shown
in Figure 9 indicate the AOB mechanism is quite effective in reducing the throughput degradation caused
by the radio interferences. In fact, while the theoretical
throughput improvement achieved by the AOB mechanism in a 3 stations scenario is 2.6% in the case of
ideal channel conditions, in our tests we measured an
improvement of the 5.7%. Similarly, the theoretical
throughput improvement achieved by the AOB mechanism in a 4 stations scenario is 4.9%, while in our
experiments we obtained a 9.4% improvement. The
reason of this better performance is that the efficiency
of the AOB scheme increases as the contention in the
network increases (as shown in Figure 9), and the
frame losses due to channel noise are treated by the
MAC protocol as normal collisions. Therefore, we
derived another very important conclusion from our
experiments: the AOB scheme shows to be useful to
reduce the probability of transmitting frames that may
get lost due to the errors induced by channel noise,
increasing the throughput in a noisy environment.
4. Conclusions and Future Work
In this paper, we presented the experiments we carried out with the implementation of an enhanced IEEE
802.11 MAC card adopting the AOB [5] backoff algorithm. Our wireless card is still fully compatible with
current implementations of the IEEE 802.11 technology because the radio part is compliant to the 802.11
standard. However, the experimental results we presented, demonstrate that the enhanced mechanism we
have implemented outperforms the standard 802.11
MAC protocol in real scenarios. We have also shown
that the advantages of this mechanism go further than
the high contention scenarios (e.g., ad hoc networks),
for which it was designed, because it is also effective
in lessening the negative impact of the external interferences, which traditionally decrease the performances
of wireless networks in any environment.
As discussed in the introduction, the wireless network interface is designed for supporting several types
of experiments with different enhanced MAC protocols. To this aim, we are working in several multiple
directions [13], [14]:
• Credit-based AOB: In recent paper [12] the
authors show that the feedback-based mechanisms
that have been proposed to maximize the IEEE
802.11 MAC protocol efficiency by guaranteeing
an optimal time-spreading of the users’ access suffer from a significant unfairness problem when employed in heterogeneous wireless networks, i.e.,
networks formed by both enhanced stations using
the modified 802.11 MAC protocol, and legacy
stations adopting the standard 802.11 MAC protocol. In particular, any contention control mecha-
nism aiming at enforcing a maximum contention
level in the network is vulnerable to the presence of
stations that are allowed to exceed this limit. For
this reason, we will experiment the implementation
of an enhanced version of AOB [12], where there is
a component capable of estimating how long the
enhanced stations further defer their frame transmissions with respect to the legacy stations, and
then exploiting this information to enable the enhanced stations to reclaim new transmission opportunities. Specifically, the enhanced stations earn
credits when releasing transmission opportunities
granted by the standard protocol. The collected
credits are a virtual currency spent by the enhanced
stations to reclaim new transmission opportunities.
• Routing: In addition to the backoff issue, there is
another problem that limits the performances of
wireless traffic in WLANs and ad hoc networks:
the routing. Indeed, each packet received from the
wireless interface must be passed up to the routing
layer (in order to discover the next hop), and further down to the same wireless interface for transferring it to the next hop. In the meantime the
SIFS time has gone and the station must run again
the backoff mechanisms in order to acquire the
channel to forwarding the packet. This adds undesirable delay and overhead at both MAC and routing layer. For this reason we aim at experimenting
mechanisms capable of executing routing operations inside the MAC, layer. The solutions we intend to experiment will be based on a next-hop address lookup in conjunction with a path strategy
as, for example, the fixed length labels architecture
as defined in [7]. Basically, the packet forwarding
protocol builds on the IEEE 802.11 DCF MAC
protocol by exploiting some additional information
delivered in the control packets (RTS/CTS) to allow the forwarding node to determine the next hop
node while contending for the channel. Moreover,
we intend to add a communication interface between the MAC layer and the routing layer, such as
to allow the routing protocol to take its routing decisions exploiting channel information.
• Cross-layering: Besides the routing, research has
shown that several mechanisms can profit from the
knowledge of some parameters that are typically
confined at the MAC layer, such as transport,
power management, cooperation, etc. Indeed, several network architecture designers foresee the access to MAC parameters for a full integration of
mechanisms traditionally working at different layers. This will be enabled by a cross-layering architecture as the one proposed by the MobileMAN
project [10]. In this architecture the shared memory
component acts as exchange area of networking information (parameters, status, etc.) for all the layers. This allows the MAC layer to distribute
“physical” information up to the higher levels, as
well as to profit from some higher layer elaborations too complex to be performed at MAC. A
typical example is the interaction between MAC,
routing and transport information for congestion
and network utilization purposes. If the transport is
aware of the links’ status, it can distinguish between congestion due to physical failures and congestion due to the amount of traffic, acting consequently. Similarly, the routing can decide different
routing paths or strategies, and the MAC can modify the distribution of some information as consequence.
5. References
[1] ANSI/IEEE Std. 802.11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications, August 1999.
[2] G. Bianchi. Performance Analysis of the IEEE 802.11
Distributed Coordination Function. IEEE Journal o n
Selected Areas in Communications, 18(9): 17871800, 2000.
[3] Xu Shugong, T. Saadawi. Does the IEEE 802.11 MAC
Protocol Work Well in Multihop Wireless Ad Hoc
Networks?, IEEE Communications Magazine,
39(6):130-137, June 2001.
[4] L. Bononi, M. Conti, L. Donatiello. Design and Performance Evaluation of a Distributed Contention
Control (DCC) Mechanism for IEEE 802.11 Wireless
Local Area Networks. Journal of Parallel and Distributed Computing, 60(4):407-430, April 2000.
L. Bononi, M. Conti, E. Gregori. Run-Time Optimization of IEEE 802.11 Wireless LANs performance. IEEE
Trans. Parallel Distrib. Syst., 15(1):66-80, January
[6] F. Calì, M. Conti, E. Gregori. Dynamic Tuning of the
IEEE 802.11 Protocol to Achieve a Theoretical
Throughput Limit. IEEE/ACM Trans. Networking,
8(6):785-799, December 2000.
[7] A. Acharya, A. Misra, S. Bansal. A label-switching
packet forwarding architecture for multi-hop wireless
LANs, in Proc. of WoWMoM 2002.
[8] R. Bernasconi, I. Defilippis, S. Giordano, A. Puiatti.
MobileMAN Wireless Network Interface,
[9] R. Bruno, M. Conti, E. Gregori. Optimization of Efficiency and Energy Consumption in p-Persistent
CSMABased Wireless LANs. IEEE Trans. Mob. Comp.,
1(1):10-31, March 2002.
[10] M. Conti, S. Giordano, G. Maselli, G. Turi. CrossLayering in Mobile Ad Hoc Network Design, IEEE
Computer, 37(2):48–51, February 2004.
[11] R. Bruno, M. Conti, E. Gregori. Optimal Capacity of
p-Persistent CSMA Protocols. IEEE Commun. Lett.,
7(3):139-141, March 2003.
[12] R. Bruno, M. Conti E. Gregori. Distributed Contention Control in Heterogeneous 802.11b WLANs, i n
Proc. of WONS 2005, Jan. 2000.
[13] R. Bernasconi, I. Defilippis, A. Puiatti. Issues and
implementation plan of an enhanced MAC, draft vers i o n
a v a i l a b l e
a t
[14] R. Bernasconi, I. Defilippis, S. Giordano. A. Puiatti.
an enhanced MAC architecture for multi-hop wireless
networks, in Proc. of PWC2003, Venice.
Experiments of Ana4: An Implementation of a 2.5 Framework for Deploying
Real Multi-hop Ad hoc and Mesh Networks
Nicolas B OULICAULT, Guillaume C HELIUS and Eric F LEURY
INSA de Lyon – France
We consider the problem of interconnecting several
hosts in a spontaneous hybrid network, i.e. an environment where wired and multi-hop wireless technologies are used. Dealing with the issues raised
by such hybrid networks and addressing all the challenges listed in MANet may require a departure from
classical solutions available in the literature. To enable a full multi-hop connectivity without raising problems/inconsistencies regarding IP compatibility, we
have proposed Ana4, a 2.5 layer architecture suitable
for hybrid networks. In this article, we present the advantages of Ana4 in the context of ad hoc and mesh
networks as well as application fields where Ana4 is
already used. We also present experimental measurements that show good performances with respect to
a full IP solution (IP routing) and similar 2.5 approaches like MPLS.
1. Introduction
Wireless communications will play a crucial role in
ambient networks that will interconnect a wide range
of equipments, from places like homes and offices to
public areas. Such ubiquitous computing will indubitably interconnect a set of heterogeneous equipments
(personal electronic devices, video entertainment, play
station, home appliances) with several heterogeneous
physical and link layer technologies, from wired to
wireless ones. In such scenarii, wireless devices may
be fixed or not and one may imagine that all devices
will take profit from a global Internet connectivity.
Wireless technologies offer open solutions to provide
mobility and services where the installation of a complex wired infrastructure is not possible. Nowadays,
performance of wireless local area networks increases
with the evolution of the normalized physical layers
(802.11b, 802.11a, 802.11g, zigbee, wimax, ...).
Multi hops wireless ad hoc networks and mesh networks offer a promising application field as main actual wireless technologies used by personal devices
operate only over short distances. Moreover, it may
not be possible to always deploy a networking infrastructure (cellular or wire cable) due to practical or cost
constraints. In order to extend the coverage of classical
wireless infrastructure-based networks, wireless multi
hop networks (or wireless ad hoc/mesh networks) have
been proposed.
In this paper, we present the implementation of a
practical architecture suitable for interconnecting devices in an hybrid network environment, where wired
and multi hop wireless technologies are used. By ad
hoc architecture, we denote a set of rules and operations dealing with addressing and routing that must be
set up for the ad hoc network to offer basic services
such as routing, IP compatibility or Internet connectivity. In particular, such an architecture must answer
the two following fundamental questions: what is an
ad hoc address? What element is identified by an ad
hoc address?
Previous works on ad hoc networks have mainly
relied on the IETF MANet working group that proposes an architecture in which the basic element is the
network interface and where an ad hoc address is an
IP address. Under IPv4, the actual philosophy is to
design/implement all MANet routing protocols at the
IP level. This architecture is not fully satisfying and
introduces several issues/inconsistencies regarding IP
compatibility. As ad hoc packets are routed using IP
addresses, there is a chicken and egg problem while
dealing with IP address auto-configuration in a multihop environment: under the MANet philosophy, in order to performe multi hop communications an IP address is mandatory bu it the same time, to obtain an
IP address from a DHCP server located several hops
away, an host need to perform a multi hop communication. Regarding this architecture, IP broadcasting in
an ad hoc network is also a hard task as the IP broadcast address must not be routed/forwarded. Moreover,
it is important to notice that, as stated in [4], a MANet
node using wireless technologies A and B (e.g. frequency A and frequency B) can communicate with
any other node possessing an interface with technology A or B. This means that the ad hoc routing must
operate over a multi-graph composed of several physical graphs and that an ad hoc node is the union of
all its interfaces involved in the ad hoc network. This
view increases the routing complexity and the number of control messages when several interfaces of a
same ad hoc host are identified by different ad hoc addresses as it is the case in the MANet philosophy. Addressing all the challenges listed in MANet [4] may
require a departure from solutions available in the literature. Implementing the ad hoc support below IP is
not new. In the past, several architectures have been
proposed: LUNAR [12] , ABR [11]. More recently, a
solution called Lilith for spontaneous networks based
on MPLS [13] has been proposed. We will see that
all these architectures lack flexibility to be completely
In this paper, we present Ana4 [1] an underlay networking mechanism useful for MANETs. Ana4 defines a new generic lightweight and efficient ad hoc architecture, which relies on the notion of ad hoc virtual
interface and logical sub-networks. A virtual interface
is a logical entity which abstracts a set of network devices into a single addressable network component. In
this paper we focus on the architecture and on the implementation and performance measurements of our
2.5 proposal for multi hop hybrid networks. The rest
of the paper is organized as follows. We first discuss
in section 2 the services that we can legitimately expect in an ad hoc network. In section 3, we present
the related works and describe our Ana4 architecture.
Then we briefly expose some advanced functionalities
of Ana4 in section 4. Section 5 is dedicated to several actual use of Ana4. In section 6 we present some
measurement experiments. Finally, we conclude this
article with section 7.
Plea for an ad hoc architecture
The fundamental service in an hybrid environment
is to allow communication between all devices of the
network, that is, to able a peer-to-peer mobile routing
capability in a purely wireless domain as stated in [4].
Since some nodes may be out of range or since some
nodes may not share the same medium (incompatible wireless devices), it is necessary to define complex
routing mechanisms that allow multi-hop routing. Inside a given hybrid network, routing mechanisms must
be implemented in order to guarantee the intranet connectivity. These mechanisms must ensure unicast and
broadcast/multicast routing capabilities. Some other
services are also required, such as support for the
TCP/IP protocol stack or Internet connectivity. In this
section, we give a brief and not exhaustive overview
of the main services that one can legitimately expect
in an ad hoc or mesh network.
Intranet connectivity. The routing paradigm is the
main factor driving the design of all networks. The
routing function problem in ad hoc network appears
as a crucial point and specific routing algorithms must
be derived for ad hoc network. Several researches are
done both on proactive and reactive approaches. It
is important to notice that, as stated in [4], a MANet
node using wireless technologies A and B (e.g. 802.11
and Bluetooth) can communicate with any other node
possessing an interface with technology A or B. This
means that the unicast routing algorithm must operate
on a multi-graph composed of several physical graphs
and that an ad hoc node is the union of all its interfaces involved in the ad hoc network. The unicast
routing must offer a global connectivity over all the
interfaces. The second important service that must be
supported in an ad hoc network is the broadcast facility. As in classical networks, a node may need to send
a message to all other nodes. This facility is used in almost all unicast routing algorithms [8, 10] developed
for MANet networks and must be supported in a efficient way at the ad hoc level.
Complete support for TCP/IP. Once the connectivity is provided, the second service that must offer an ad
hoc network is the TCP/IP one as the whole Internet
relies on these protocols. Every node must be able to
behave as if it belongs to a standard IP network, that is
in ”an interoperable inter-networking capability over
a heterogeneous networking infrastructure”. Moreover, a partial support or compatibility with IP is an
unsatisfying approach. Based on these trivial remarks,
we must focus on the consequences they imply. First,
IP defines a set of addressing as well as address-related
routing rules. For example, it defines the notion of IP
networks and IP sub-networks. Routing and accessibility directives are associated to these notions. IP also
proposes broadcast notions and rules. For example, a
packet directed to the address is received by all nodes connected to the local link of the
source and is not supposed to be forwarded. Numerous protocols and applications rely on IP standards. If
we desire a complete compatibility with existing networking environments, the ad hoc architecture must
be fully compatible with IP. Secondly, several autoconfiguration mechanisms are proposed in association
to IP. In IPv4, the DHCP protocol allows an host to retrieve its IP address from a server. If these services are
powerful ones in wired networks, their importance is
also obvious in ad hoc networks.
Internet Connectivity. In an hybrid context, it is important to offer a global connectivity. The notion of
global connectivity is more general than the care of
address feature present in IP mobility or the delivery
of a global IPv6 address. Offering a global connectivity to the Internet is providing a service continuum.
This means for example that an ad hoc node must be
able to receive its favorite net-radio multicasted from
a server which is not localized within the ad hoc network. As stated above, the multicast protocol used inside the MANet will be specific and the global connectivity service will perform the gateway operations with
Other features. We may want to be able to switch
from one physical radio interface to another for power
consumption reason for example, without interrupting
its sessions or having to deal with any IP mechanism.
It means that a micro-mobility (or vertical mobility)
support must be implemented inside an ad hoc node
so that changing its ad hoc communicating interface
will not lead to a change in the IP address used by the
host. An other very important feature is the easiness
of implementation. Turning a node (PC, PDA...) into
an ad hoc node should not require any kernel modification nor any driver modification. Last but not least,
the scalability requirement may be an important factor.
One can imagine than an ad hoc network may support
tens to hundreds of mobile nodes but must also be scalable to a higher factor if this kind of networks becomes
really pervasive [7].
3. The Ana4 Architecture
Based on the preceding remarks, it seems important
to provide a more fine grain layering that includes an
ad hoc level as compared to the pure IP layering. We
can locate it at level 2.5, i.e., between level 2 (MAC)
and level 3 (IP). Locating the ad hoc level at layer 3
induces several problems with the IP protocol as it is
currently known. We argue that in order to be able to
deploy ad hoc networks in a very friendly way, a full IP
stack support must be provided which also implies to
take into account the legacy of already deployed applications/configurations. Likewise, locating the ad hoc
level at layer 2 like for the LUNAR [12] architecture
induces several issues concerning multiple interfaces
support and hardly enables support for micro-mobility.
For instance, LUNAR does not address the support of
multiple interfaces. Moreover, LUNAR is bound to
a specific routing scheme and has been designed for
small networks of around 10 to 15 nodes.
Locating the ad hoc layer at level 2.5 is an interesting alternative as it can both enable a full support
of IP and provide the opportunity to handle several interfaces. A first step towards this direction was made
with an implementation of ABR [11]. However, ABR
relies on the IP addressing which leads to the already
described auto-configuration issue and, as LUNAR,
is bound to a specific routing scheme. Recently, a
work posterior to Ana4, Lilith [13] was designed to
use MPLS for ad hoc routing. However, this solution is
also not satisfying as it does not support multiple interfaces or vertical handoffs, the handling of logical subnetworks or auto-configuration mechanisms and since
it is binded to a on purpose proactive like routing pro-
Inter-node architecture
We propose an architecture breaking up an ad hoc
network into three levels of abstraction: the hardware
level, the ad hoc level and the IP level. If the first
one relies on a physical reality, interface communication compatibility, the two others are purely theoritical
views. The base element of the first level is the wireless interface whereas for the two others, it is the ad
hoc node. Our definition of an ad hoc node follows the
one proposed by MANet in [4].
Hardware level. The hardware level is the set of the
different hardware networks. A hardware network is
the gathering of all interfaces that are physically able
to communicate with each others. At this level, the
notion of communication ability is related to link layer
device compatibility and not to effective communication possibility. In a hardware network, hardware addresses (e.g., MAC) identify interfaces.
Ad hoc level. The ad hoc level defines the ad hoc
network. An ad hoc network is the combination of
all hardware networks. At this level, the base element
is no more the interface but the ad hoc node. We do
not make distinction between interfaces but only see
nodes with a single interface, the ad hoc virtual interface connected to the ad hoc network. A unique ad
hoc device is abstracted from all wireless and wired
devices. In an ad hoc network, an element is named
using a unique node identifier, also called an ad hoc
address. Multi-hop communication is available. One
node may send packets to a node distant from several
hops. Packets are commuted from ad hoc nodes to
ad hoc nodes depending on the ad hoc addresses of
destinations. Broadcast and multicast mechanisms are
also available. While commuted, a packet may transit through any underlying hardware network and join
the destination through any of its hardware interfaces.
The followed path is determined by the commutation
or routing protocol. This routing protocol is independent of the architecture.
IP level. At this level, an ad hoc network is seen as
an Ethernet bus (more precisely as a switched Ether-
net link): the IP abstracted network. An ad hoc node
of the ad hoc level is looked upon as a single and classical Ethernet interface: the virtual ad hoc interface. In
other terms, a node with several physical devices only
owns one interface from the IP view. All the commutation work performed at the ad hoc level is transparent
to IP.
This architecture allows a complete compatibility
with IP. For example, locally broadcasted packets
reach all nodes of the ad hoc network without being IP
routed. Auto-configuration becomes straight-forward.
Retrieving an address at multiple hops through the
DHCP protocol is possible since ad hoc nodes are
reachable even if not IP configured. As the commutation is performed at the ad hoc level, no IP address
is needed to communicate with other ad hoc nodes.
More globally, all the IP world behaves as it does with
an Ethernet link.
Intra-node architecture
The role of a virtual ad hoc interface, illustrated in
figure 1, is to hide the different physical devices and
hardware networks; it provides the illusion of a single
virtual network. At the ad hoc level, this virtual network is a wireless multi-hop network; at the IP level, it
is a switched Ethernet link. A powerful characteristic
of this architecture is to allow an host to use a device
simultaneously in ad hoc and in classical modes. Suppose that a physical device handled by a virtual ad hoc
interface is also configured as an Internet device. From
the IP view, the mobile hosts two distinct interfaces.
IP networking is performed over these two interfaces
without interference.
The virtual interface. For upper layers, the virtual
interface acts as a classical interface. For example, it
is declared as an Internet device to the IP layer. IP outputs packets directed to ad hoc nodes through the virtual interface. For the under layer, i.e. the link layer,
the virtual interface acts as an upper layer protocol.
Upon reception of a packet that has transited through
the ad hoc network, the packet is given to the virtual interface. This particular architecture is interesting since
it does not require any modifications neither in device
drivers nor in the TCP/IP stack.
global architecture. Any MANet protocol may potentially be used. Its role is to compute or discover routes
and to configure commutation tables in ad hoc interfaces. Very few work is required to adapt a MANet
routing algorithm to our architecture. Basically, node
identifiers have to be changed from IP to ad hoc addresses.
ad hoc network
from IP
to IP
Level 4
IP address
Level 3
ad hoc
ad hoc virtual
Level 2
ad hoc address
from / to interfaces
Figure 1. The virtual interface
Naming ad hoc interfaces. Introducing a logical
network and logical interfaces require the introduction
of a corresponding logical naming process. Virtual
interfaces are addressed using ad hoc addresses. An
ad hoc address is composed of two fields (Fig 2): the
network identifier field, Net Id, and the node identifier field, Node Id. The last field must ensure the
uniqueness of the ad hoc address. It may be configured by hand, chosen using a MAC address or using
a statistically unique and cryptographically verifiable
(SUCV) identifier as presented in [9]. SUCV identifiers allow authentication of ad hoc nodes in the network.
Advanced functionalities
As we have already said, our architecture intrinsically provides a multiple interface support and allows
the use of different network protocols over the ad hoc
network. Ana4 is an architecture that allows a complete support for IPv4, including auto-configuration
mechanisms. It also provides a complete connectivity
with the Internet, in regards to routing but also multicast and other services mechanisms. Moreover, Ana4
handles problems related to scalability and network
partitioning. Trying to keep a totally flat topology may
induce too long flooding delays or huge routing tables. The network identifier field of an ad hoc address
allows the setup of sub-ad hoc-networks through the
introduction of communication policies between ad
hoc nodes with different network identifiers. Ad hoc
networks can be split in several logical subnetworks
which will appear to IP as different virtual Ethernet
links. For more information on Ana4, its functioning
or its advanced functionalities, please refer to [2].
Node Id
Figure 2. Structure of an ad hoc address
MAC address
The role of the virtual interface is to commute packets between different devices and upper layer protocols. Upon reception of a packet, the interface decides
whether it has to emit the packet, through which interface and to which nodes, and whether it has to forward
it to upper layers. A virtual interface owns a commutation table. This table is managed by an application
level routing protocol such as the ones studied in the
MANet group (OLSR [8], AODV [10] or DSR [6] for
example). The routing protocol is independent of the
Deployment of Ana4
Ana4 is currently used in 4 different test beds, two
of them in partnership with private companies.
Classical mesh network research test bed. The
first “simple” scenario is to use Ana4 to set up a wireless mesh network test bed using an ad hoc routing
protocol such as OLSR. By mesh network, we assume
a network composed of some fixed and mobile nodes,
wireless and wire links and that handles many-to-many
connections and is capable of dynamically updating
and optimizing these connections. The dynamic management of complex routing information that includes
information about external networks (e.g. the whole
wide Internet and the gateways to it), is one important
key feature inside our research lab mesh network. By
deploying 10 shuttle PCs in the lab, some of them connected to the Internet via a wired link, and by configuring several laptops we are able to use this hybrid network as a regular wireless office covering network. We
can come up with relevant real-world scenarios where
hybrids between ”static” and ”ad-hoc” networks offer
some clear advantages – networks where a number of
static nodes form the matrix (the substrate) in which
other nodes appear, roam, and disappear. In this mesh
network, auto-configuration of mobiles is performed
using DHCP.
Heterogeneous seismic sensor network. The
IHR [5] project is the second test bed using Ana4. The
ambitious goal of the IHR project is to develop a new
seismic tool that would allow seismic investigation
at scales comprised between one kilometer and few
hundreds of meter. The geological targets are those
potentially dangerous areas: fault zone, volcanoes,
land-slides, valley with site-effect. A new equipment
has been recently delivered. The new seismic network
consists of thirty nine channels data-loggers equipped
with six vertical sensors plus one component sensor.
These spider-like mesh networks are connected to
each others by network links (wire when possible
and/or wireless) that allow a limited crew to control
and tune the 270 channels. The deployed seismic
network is heterogeneous and scientists on the field
encounter several pure networking problems when
they try to manage this spider network. How to
configure the wireless network, the wire backbone
network ? How to deal with IP sub networks... Ana4
offers a simple solution. Due to the Ana4 support
for multiple interfaces, the ad hoc network is spread
over both the wireless and wire links, hiding the
complex and heterogeneous topology to the end user.
Auto-configuration is also made straight-forward as
the retrieval of an IP address is possible even if nodes
are several hops away from the DHCP server. In
order to manage the seismic network, scientists also
perform broadcast operations in order to start/stop all
data loggers. Once more, there is no need to modify
all specific application dedicated to the control of the
seismic sensors as Ana4 offers a real broadcast to the
IP layer even if packets are doing multi hops at the ad
hoc layer.
Video Billboard. Those who communicate messages are always in search of attractive ways to carry
their information more effectively. When placing
outdoor advertising campaigns, 30-second television
commercials on large screens (14.69ft x 11.02ft), displaying advertisements, video footage and general information on buildings and walls in busy metro, market, restaurant and/or nightlife districts, the main problem is to carry the messages and update the information. The key idea was to use Ana4 to build a “video
billboard mesh network” in order to perform content
delivery and network management without deploying
a wired network. It’s always more easy to get a power
supply on outside building wall than a RJ45 like plug.
This project is in partnership with the Embedia company. Embedia is an interactive communications solutions provider. Embedia creates an innovative stateof-the-art link between businesses and their consumers
and its patented solution delivers interactive multimedia content directly to end-user devices.
Inventory control and localization. In many industrial contexts (logistics, objects and people monitoring,
security...) it may be essential to localize precisely objects or people in real time, whether they are situated
indoor or outdoor. In this goal, Kadya is working on
providing a localization solution based upon cheap and
light emitting radio tags, monitored by a set of listening stations. Each station collects data emitted by the
tags through a proprietary radio protocol, and reemits
them through a mesh network to a server which analyzes the radio data to calculate the position of each
radio tag. Ana4 is used to provide the illusion of a
one hop IP network, allowing use of DHCP and broadcast operations for the listening stations. Moreover, the
system is easily extensible as any additional station can
be placed in the mesh network without the need for any
configuration as the multi-hop connectivity is achieved
by Ana4 and the IP configuration by DHCP. Furthermore, Ana4 allows the use of IP broadcasting over the
network to monitor the system.
6. Performance
We have implemented Ana4 as a kernel module and
the code has been released1 . The first implementation
was done under Linux (PC and PDA) and Windows
XP. We provide more details on the implementation
in [3] and we summarize below some performance results based on analyses and measures in an experimental network of Ana4 nodes.
Figure 3. Effective bandwidth ratios on the y axe as a function of the transmitted packet
size (in bytes) on the x-axe.
First, we show the theoretical overhead of Ana4.
The Ana4 header introduces an overhead of 160 bits
in all packets transiting in the ad hoc network, reducing the performance the network may achieve. However, this overhead is not significant. First, it does not
much reduce the volume of data a packet may include.
160 bits is less than 0.8% of a 802.11 link transfer
unit (TU: 2312 bytes). Second, this overhead does not
much lower the useful bandwidth of the medium. It
only takes 14us to transfer 160 bits over a link with
a 11Mbits/s throughput. In the 802.11 technology,
this delay is 8 times smaller than the delay awaited
before accessing the medium (DIFS= 128us). Figure 3 presents some theoretical bandwidth ratios. The
medium is a 802.11 link with 11Mbits/s throughput.
The x-axis is the size of data in the exchanged packets. The y-axis is the ratio between the effective bandwidths with and without Ana4. The two plots correspond to a classical RTC-CTS-Data-ACK transfer scheme and a Data-ACK one. In the worst case,
packets with only 1 byte of data, the overhead is very
low, only 7% loss of the effective bandwidth.
To evaluate our Ana4 prototype, we measured its
performance in a multi hop network composed of a 4
nodes chain connected via 802.11b wireless or wire
links (Fig 4). The goal is to measure the round trip
time (ping) and the throughput (netperf) between
nodes at distance 1, 2 and 3 hops (i.e., between nodes:
A − B, A − C and A − D). The comparison was done
between: (i) a pure static IP route scheme which may
be considered as the reference when implementing ad
hoc network at layer 3, (ii) MPLS commutation with
MPLS linux2 which may also be considered as a reference when implementing an ad hoc solution at layer
2.5 and (iii) our Ana4 implementation.
1 hop: A − B
2 hop: A − C
3 hop: A − D
(ii) MPLS
(iii) Ana4
Table 1. Latency comparison in a wire environment. Average on 100 ping measures with
a 64 byte data payload (on top of IP).
Table 1 presents the result of the latency measurements with wire links. The reasons why we used
wire links for this measurements is to minimize the
medium transmission time in order to isolate the overhead caused by the Ana4 or MPLS layer. We can observe that the overhead introduce by Ana4 is very low.
In this overhead, one part is caused by memory operations (introduction and deletion of the Ana4 header).
This part can be observed in the one hop measurement.
As we can see, this setup cost is equivalent for Ana4
(0.115ms) and MPLS (0.114ms) as both protocols perform equivalent operations with their respective headers. The rest of overhead is due to computations (routing table interrogation), at each hops. For this last part,
MPLS achieves a better performance (0.092ms/hop for
MPLS vs 0.105ms/hop for Ana4) even if the difference
is quite negligible (0.013ms). The reason for this is
that the MPLS header (32 bits per label pushed in the
MPLS label stack) is smaller than the Ana4 one and
that the MPLS code is highly profiled and designed
for high performances which was not the case with the
Ana4 prototype. We are confident on the fact that we
could achieve similar results as MPLS with a profiled
Ana4 code.
Figure 4. Experimental platform.
(i) static IP
1 hop: A − B
2 hop: A − C
3 hop: A − D
(i) static IP
(ii) MPLS
(iii) Ana4
2.03 Mb/s
Table 2. Throughput comparison (netperf)
in a wireless environment.
Table 2 presents the result of bandwidth performances in a wireless environment. The bandwidth was
measured using the netperf software. In the worst case,
1 hop, Ana4 induces a bandwidth lost of 3%. This
time, the overhead decreases with the number of hops.
The reason for this is that the computation overhead at
each hop is recovered by the communications during
the transfer. As a consequence, the highest the number
of hops, the most negligible it becomes. As for the latency, the difference between MPLS and Ana4 is due
to the headers size and the code profiling.
7. Conclusion
We have presented Ana4, a practical architecture
suitable for interconnecting devices in hybrid network environments. Our goal is to provide a generic
lightweight and efficient ad hoc architecture, which
relies on the notion of ad hoc virtual interfaces and
logical sub-networks. Ananas defines an ad hoc network by introducing three levels of abstraction: the
hardware level , the ad hoc level 2.5 and the IP level
3. By designing an ad hoc level at layer 2.5, actual
ad hoc routing protocols can be implemented and advanced features like sub-networking, vertical handover
or auto-configuration can be offered.
Our first implementation under linux and windows
XP shows good performances compared with traditional IP routing and/or MPLS commutation. We observe only a small degradation for the latency and a
very small overhead for the throughput when using
the Ana4 architecture. However, and as stated before,
Ana4 offers advanced features like sub-networking,
auto-configuration, vertical handover or all ad hoc
node broadcast...
Several possible extensions for this work are open
to investigation: taking into account the energy consumption of interfaces to switch from one to another
under the virtual ad hoc interface in order to provide
the best ratio between energy and throughput for a
given class of traffic; auto-organizing the network using the ad hoc sub-network possibilities. Last but not
least, we still work on the performance experiments,
on the Ana4 code profiling and on the deployment of
Ana4 in several scientific or industrial real test beds as
mentioned in this paper.
[1] G. Chelius and E. Fleury. Ananas : A local area ad
hoc network architectural scheme. In MWCN 2002,
Stockholm, Sweden, Sept 2002. IEEE.
[2] G. Chelius and E. Fleury. Ananas : A New Adhoc
Network Architectural Scheme. Research Report RR4354, INRIA, March 2002.
[3] G. Chelius and E. Fleury. Ananas : A new adhoc network architectural scheme. RR 4354, INRIA, 2002.
[4] S. Corson and J. Macker. Mobile ad hoc networking
(MANET): Routing protocol performance issues and
evaluation considerations. IETF RFC 2501, 1999.
[5] O. Coutant, F. Doré, J. Fels, D. Brunel, M. Dietrich,
F. Brenguier, and S. Judenherc. The high resolution
seismic imaging (ihr) network, a new tool for seismic
investigations at hectometric scales. Geophysical Research Abstracts, 7, 2005.
[6] Y. Hu, J. Jetcheva, D. Johnson, and D. Maltz. The
dynamic source routing protocol for mobile ad hoc
networks (DSR). Internet Draft, 2001.
[7] J.-P. Hubaux, T. Gross, J.-Y. Le Boudec, and M. Vetterly. Towards self-organized mobile ad hoc networks
: the terminode project. IEEE Communications Magazine - Special Issue on Telecomuunications Networking at the Start of the 21st Century, January 2001.
[8] P. Jacquet, P. Muhlethaler, A. Qayyum, A. Laouiti,
L. Viennot, and T. Clausen. Optimized link state routing protocol. Internet Draft), 2001.
[9] G. Montenegro and C. Castelluccia. Statistically
Unique and Cryptographically Verifiable Identifiers
and Addresses. In Proceedings of ISOC NDSS02, San
Diego, February 2002.
[10] C. Perkins, E. Royer, and S. Das. Ad hoc on-demand
distance vector (AODV) routing. Internet Draft (work
in progress), November 2001.
[11] C.-K. Toh. Ad Hoc Mobile Wireless Networks. Prentice hall, 2002.
[12] C. Tschudin, R. Gold, O. Rensfelt, and O. Wibling.
Lunar: a lightweight underlay network ad-hoc routing protocol and implementation. In NEW2AN’04, St.
Petersburg, February 2004.
[13] V. Untz, M. Heusse, F. Rousseau, and A. Duda. On
demand label switching for spontaneous edge networks. In Workshop on Future Directions in Network Architecture (FDNA), Oregon, USA, September
GNU/Linux Implementation of a Position-based Routing Protocol
Marc Heissenbüttel, Torsten Braun, Tobias Roth, Thomas Bernoulli
Institute of Computer Science and Applied Mathematics
University of Bern
3012 Bern, Switzerland
{heissen, braun, roth, bernoull}
The Beacon-Less Routing protocol (BLR) is a positionbased routing protocol for mobile ad-hoc networks that
makes use of location information to reduce routing overhead. However, unlike other position-based routing protocols, BLR does not require nodes to periodically broadcast hello messages and thus avoids drawbacks such as
extensive use of scarce battery-power, interferences with
regular data transmission, and performance degradation.
In this paper, we describe the implementation of BLR on
a GNU/Linux platform comprising laptops equipped with
802.11b WLAN cards and GPS receivers. We present results of BLR’s performance obtained from laboratory experiments, which were conducted to validate the implementation for future planned outdoor experiments.
1 Introduction
In position-based routing protocols forwarding decisions
are solely based on location information. Each node is
aware of its own position, e.g., through GPS, and of its
immediate one-hop neighbors by the periodical broadcast
of hello messages. Additionally, a location service is required that allows determining the position of the destination node, e.g., GLS [1]. Each node simply forwards a
packet to a neighbor which is closer to the destination until the packet eventually arrives at the destination. Many
position-based routing protocols have been proposed such
as GFG [2], GPSR [3], GOAFR [4]. An overview can be
found in [5]. A major drawback of those protocols is the
proactive transmission of hello messages which uses scarce
network resources such as battery power and bandwidth.
Recently, the Beacon-Less Routing protocol BLR was proposed in [6] based on a new routing paradigm enabled by
the broadcast property of the wireless propagation medium.
Unlike all other routing protocols, forwarding decisions are
not taken at the sender of a packet, but in a completely dis-
tributed manner at the receivers. A sender does not have
to be aware of its neighbors and consequently nodes do not
have to proactively transmit hello messages (beacons) as in
other position-based protocols, which also saves scarce network resources like battery energy and bandwidth. The performance and the behavior of BLR was studied analytically
and by simulations in [6]. Results show that BLR provides
efficient and robust routing in highly dynamic ad-hoc networks and is immune to topology changes. Therefore, BLR
is especially suited for vehicular and sensor networks with
frequently changing topologies. The promising results are
the motivation to go a step further and implement the protocol in a real testbed. We developed a prototype system
and conducted measurements to obtain more insight on the
protocol’s performance and behavior in a real world environment. In Section 2, we briefly review the BLR protocol
and describe its main features. Afterwards, the implementation on a GNU/Linux platform is presented in Section 3
and encountered real world challenges are discussed in Section 4. Section 5 describes the experimental setup and provides measurement results. Finally, Section 6 concludes the
2 Beacon-Less Routing Protocol (BLR)
Unlike other position-based routing protocols, BLR does
not require the periodic broadcast of hello packets. BLR
selects a forwarding node in a distributed manner among
all its neighboring nodes without knowing the existence or
positions of neighbor nodes. BLR has three main modes of
operation; greedy mode, backup mode, and unicast mode.
2.1 Greedy mode
Packets are routed in greedy mode whenever possible because only in this mode BLR is really stateless and does not
require the transmission of hello packets, i.e., nodes are normally not aware of any neighboring nodes. Therefore, when
a node has to send a packet it simply broadcasts the packet.
Consequently, all neighbors receive the broadcast packet.
The protocol ensures that just one of the receiving nodes relays the packet further. This is accomplished by different
forwarding delays and restricting the nodes that are allowed
to forward the packet to a certain area, called forwarding
area. Nodes within this area can mutually receive each others transmissions. For the forwarding area BLR uses a circle with diameter r relative to the forwarding node S in the
direction of the final destination D as depicted in Fig. 1. A
Forw arding Area
2.2 Backup Mode
If no node is located within the forwarding area, greedy
routing fails. This is detected if a node does not overhear
a further rebroadcast within M ax Delay of its previously
broadcasted packet. This node forwards the packet via unicast further in backup mode. Therefore, the node broadcasts
a request for a beacon packet. All neighbors that receive this
packet reply with a beacon indicating their positions. The
packet is then forwarded to the replying node that is closest
to the destination. If none of the neighbors is closer to the
destination than the requesting node, the packet is routed
according to the face routing algorithm based on the ”righthand” rule, a concept known for traversing mazes, on the
faces of a locally extracted planar subgraph, see for example GOAFR [4] for more details. As soon as the packet arrives at a node closer to the destination than where it entered
backup mode, the packet switches back to greedy mode.
2.3 Unicast Mode
Figure 1. Forwarding Area with potential forwarders A and B
receiving node can determine if it is within the forwarding
area from its own position and the positions of the destination D and the previous node S. Both positions of S and
D are stored in the packet header. Nodes in the forwarding area are called potential forwarders, e.g., A and B in
Fig 1. Potential forwarders calculate a Dynamic Forwarding Delay (DFD) in the interval [0, M ax Delay] depending
on their position relative to the previous and the destination
node. The DFD is calculated by (1) with r as the transmission radius of a node, p the node’s progress towards the
destination, and M ax Delay as a system parameter. Nodes
outside the forwarding area simply drop the packet (node
Add delay = M ax Delay ·
According to this DFD function, the node with the most
progress (e.g., node B), i.e., closest to the destination, calculates the shortest Add Delay and thus rebroadcasts the
packet first. The other potential forwarders (e.g., node A)
overhear this further relaying and cancel their scheduled
transmissions of the same packet. The rebroadcast packet
is also received by the previous transmitting node and acknowledges the successful reception at another node. Simultaneously, the neighbors of the rebroadcasting nodes
also received the packet and they determine if they are
within the forwarding area relative to node B and destination D. Potential forwarders calculate an Add Delay and
compete to rebroadcast the packet again.
Routing in greedy mode makes BLR susceptible to
packet duplication as data packets are broadcast over multiple hops. Packet duplication occurs for each node in the
forwarding area, which does not detect that a packet was
already rebroadcast. In reality, there are many reasons that
prevent nodes from successfully receiving the rebroadcast
packets such as irregular transmission ranges, obstacles, and
simultaneously on-going transmissions in the vicinity. BLR
implements the unicast mode to minimize the number of
duplicated packets. After a node has detected that another
node has rebroadcast the packet, it is also aware of the forwarding node’s position. Thus, the node may send the subsequent packets to the same destination via unicast to the
node which relayed the broadcast packet. Due to the mobility of the nodes, nodes located at a better position may
enter into the node’s transmission range. In order to be
able to detect these new nodes, a packet is broadcast in
greedy mode after a certain time again such that potential
forwarders compete to rebroadcast the packet.
3 Implementation
3.1 Overview
The target platform of the implementation is
We used Gentoo Linux [7], although
any other GNU/Linux distribution based on Linux 2.6 will
work for our implementation. We integrated BLR within
the protocol stack as depicted in Fig. 2, i.e., between the IP
and the link layer. Therefore, it is transparent to the upper
layers and applications. Consequently any application such
as HTTP, ssh, ping and also ICMP can run unmodified.
The BLR protocol was however implemented in the user
space of Linux due to simplicity reasons. Therefore, outgoing packets (solid line) have have to be intercepted and
processed accordingly before being passed to the wireless
network adapter. More specifically, we introduced a virtual
interface tun0 provided by the tuntap [8] device. A
new route that redirects all traffic to the BLR network
(private destination IP-Addresses through
tun0 is added to the system routing table. Consequently,
Internet traffic is not affected by the BLR application
and routed as normal directly to the 802.11 interface.
By listening on tun0, the BLR application can catch
all traffic sent to the BLR network and inserts the BLR
header and updates the IP header. Afterwards, packets
are sent via the pf_packet facility, which allows the
sending of Ethernet and IP packets directly to the 802.11
network adapter. Incoming packets (dashed line) are passed
User Space
tun0-I f.
has to be blocked somehow. This is achieved by deploying the IPtables [9] packet filter right after the pf_packet
facility. This filter blocks all incoming traffic that has the
protocol number of BLR set in the IP header. For broadcast
packets, this blocking would not be necessary since the kernel simply drops broadcast traffic with a protocol number
for which there is no open socket. However, when the kernel receives unicast traffic with an unknown protocol number, it sends an ICMP destination unreachable message back
to the sender, which has to be avoided.
3.2 BLR Application
The BLR application is split into three separate
processes as depicted in Fig. 3. The main process
receives/sends the packet from/to the localhost/network,
transforms and updates headers, calculates the Add Delay,
and manages packet timeouts, unicast route information, as
well as a list of duplicate packet IDs. The GPS process
is connected to an external GPS device and provides location information. The sendqueue process receives outgoing
packets together with the calculated Add Delay from the
main process and sends the packet after the indicated delay.
If the main processes receives a packet from pf_packet,
it calls the sendqueue process in order to determine if the
packet is queued for transmission. If so, another node forwarded the packet first and the sendqueue process can remove the packet from the queue.
The size of the BLR header is 32 bytes and has the following fields.
• Packet type (1 byte): Data, Location request, LocationReply, Request for beacon, Beacon.
Figure 2. Implementation of BLR in the protocol stack
• Original protocol (1 byte): Protocol number of TCP,
UDP, ICMP, etc. This corresponds to the protocol field
in the IP header, because the BLR header is inserted
between the IP and transport layer header.
over pf_packet to the BLR application and are either
forwarded to the next hop or passed to localhost, depending
on the destination address in the IP header. When packets
are forwarded, the BLR application only updates the BLR
header and additionally delays the packets by the newly
calculated Add Delay before the packets are passed again
via pf_packet to the network adapter. On the other hand,
when the packet is destined for this host, the BLR header
is stripped off and the IP header modifications done by the
BLR application at the sender are reversed. Afterwards, the
packet is forwarded through the tun0 to the application.
A problem occurs because pf_packet actually creates
a copy of all incoming packets. One copy is passed to the
BLR application, while the original packet is passed to the
kernel and from there to the application. The original packet
• Sequence number (2 bytes): This number together
with the source address allows to unambiguously identify a packet.
• Backup distance (4 bytes): This field is used to indicate the distance to the destination from where greedy
routing failed, which is required in order to determine
when to switch back to greedy routing.
• Position information (8 bytes each): Position of the
previous, source, and destination node. The previous
and destination node positions are required to calculate
Add Delay. The source node’s position is used to update location information at the destination in case of
bidirectional traffic.
In the following, we describe in more detail the processes
and how packets are handled.
• The packet length fields needs to be increased by the
size of the BLR header.
tun0-I f.
tun0-I f.
Position/ Time
• The protocol field is changed to 254, which we used
for the BLR protocol. The original protocol number is
stored in the BLR header.
• Finally, the header checksum needs to be recalculated.
Figure 3. Running processes in the BLR application
GPS Process
The GPS process is connected to an external GPS device, which it polls for changes in location information. It
parses the GPS data and passes position updates to the main
process. The connection is established through an RS-232
interface and the GPS information is transferred with the
NMEA-0183 protocol [10] from the GPS receiver to the
Sendqueue Process
The sendqueue process is responsible for queuing the packets according to their respective dynamic forwarding delay.
It receives packet/delay tuples from the main process and
maintains an ordered list of all pending packets. When
the associated timer expires, the packet will be sent to
pf_packet. The sendqueue is implemented separately
from the main process since it its only tasks are to manage packet delays and to queue and send packets. The
sendqueue further handles the deletion of packets from the
queue, whenever the main process detects that another node
has already forwarded a pending packet.
Main process
This is by far the most complex process and is responsible
for switching packets between components, management
and coordination of the other components, execution of the
BLR functions, etc. When the main process receives an
IP packet through tun0, it inserts the BLR header and updates the IP header. Changes are necessary in four IP header
• The source address needs to be changed from the IP
address of the tun0 interface to the IP address of the
outgoing interface.
Furthermore, the main process calculates Add Delay and
forwards this along with the packet to the sendqueue
process. It also maintains a hosttable to store information about known destinations, namely their most recent
positions and the next hop to reach them if unicast mode
is used. The packetlist caches packets that have been
sent and have not yet been acknowledged, together with a
timeout value for each packet. packetlist also handles
retransmissions in case of timeouts. Whenever BLR has to
switch to backup mode, it takes some time until the next
hop is determined due to the sending of the beacon request
packet and the time until the beacons from the neighbors
are received. The backupqueue caches outgoing packets that have to wait to be forwarded until the backup mode
setup is completed.
4 Challenges
In this section, we briefly review the main challenges we
faced during the implementation in the Linux testbed as opposed to the previous implementation in the simulator.
4.1 Location Service
Location services that provide the position of nodes is a
research aspect in itself and several solutions have already
been proposed (see [5] for an overview). Therefore, a common assumption of most position-based routing protocols is
that the position of the destination is somehow known. In
the network simulator, it can be implicitly assumed that this
position information is available. In reality, we have to implement a mechanism that provides the position. As it was
not our goal to implement a fully functional location service
and it would not be appropriate for a small testbed with a
few laptops, we chose to implement a simple request-reply
mechanism based on flooding. The request and the reply
with the geographical position are piggybacked on data traffic whenever possible. In case of unidirectional traffic, position information is invalidated periodically, and the source
broadcasts a new location request. In case of bidirectional
traffic, or simply if TCP is used, destination locations are
not invalidated, but the position can be simply extracted
from packets returning from the destination, namely from
the source field in the BLR header. Thus, the overhead can
be reduced to the initial flooding of one location request
packet in case of bidirectional traffic.
4.2 Duplicated Packets
4.5 Interrupt granularity
The objective of the unicast mode is to reduce the number of duplicated packets. However, still transmissions are
broadcast over intermediate hops. In ideal conditions of a
network simulator, radio propagation is modeled by simple isotropic transmission ranges. In reality however, we
observed many duplicated packets due to irregular transmission ranges. Therefore, we additionally implemented
a filtering mechanism. Each node compares the uniquely
identifying source address and sequence number of a packet
against a table containing the recently received and also
overheard packets. If this packet is a duplicate of a previously received or overheard packet, the node broadcasts
a control packet suppressing the further forwarding of that
duplicated packet by its neighbors. Therefore, the duplicated packet can be again disposed.
The M ax Delay can be chosen in the order of some
milliseconds based on the experienced network simulator
results. Basically, M ax Delay indicates the range over
which potential forwarders schedule their retransmissions.
The Linux kernel has a limitation that severely affects
the possible value of M ax Delay, namely the granularity of the timer interrupts. This granularity is defined by
a compile-time kernel constant called HZ. On Linux kernels
2.6 or newer, this constant is set to 1000 resulting in timer
interrupts every 1 millisecond. (In kernel 2.4 and older,
the HZ was set to 100). This means that the select()
system call returns at 1 millisecond intervals only. Consequently the granularity of Add Delay is also only 1 millisecond. Therefore, a rather long M ax Delay has to be
chosen to reduce the risk that all nodes transmit simultaneously and limit the usefulness of the DFD concept. In [6],
it was proposed to set M ax Delay = 2 ms based on simulation results, which is definitely too short for the Linux
implementation. However, the longer M ax Delay also increases the end-to-end delay. While possible in theory, a
further increase of the HZ value is not yet completely supported by the Linux kernel. Even if possible, increasing
HZ also increases the overall timer overhead, because more
timer interrupts are generated. This may not be an issue
for our testbed where no other applications are running, but
definitely it will be an issue for small mobile devices with
limited computation resources.
4.3 IP Fragmentation
The BLR header is part of the IP payload in the current
implementation. Thus, if fragmentation occurs, only the
first IP fragment will contain the BLR header. The header
is however required to route packets by BLR. Subsequent
fragments will not contain the BLR header and will simply be dropped, because nodes do not know how to process
them. Therefore, IP fragmentation has to be avoided. To
achieve this, the MTU of the virtual tunnel interface is decreased by the size of the BLR header, which is inserted before Transport layer header, in order to avoid fragmentation
at the source node. Additionally the DF (Don’t Fragment)
bit is set in the IP header such that intermediate nodes do
not fragment the packet. PMTU (Path MTU) discovery is
used to handle links where the standard MTU is too large.
4.4 MAC layer control
If a unicast packet is not acknowledged, the 802.11 MAC
layer retransmits a packet up to seven times before giving
up. In the network simulator implementation, the MAC
layer can signal a failed transmission to the upper layer,
which in turn selects another next hop and passes the packet
again to the MAC layer. This mechanism is also applied by
BLR [6] and GPSR [3]. Without this optimization, many
unicast packets would be dropped due to unreachable neighbors, and recovery is left to TCP or the application. This
severely decreases network performance as retransmissions
are end-to-end and not link retransmissions. In a Linux
implementation of BLR with WLAN cards however, the
MAC protocol is largely implemented in the firmware of the
802.11b card, which makes accessing the mentioned functions in today’s card nearly impossible.
5 Experiments
5.1 Equipment and Configuration
The testbed consists of 5 laptop computers running
Linux 2.6. Each laptop is equipped with an IEEE 802.11b
WLAN cards. The cards are configured to run in ad-hoc
mode without RTS/CTS, i.e., the DCF of 802.11b is used,
and the data rate is set to 2 Mbps. The hardware equipment
is heterogenous, i.e., the laptops are from different manufacturers. The same applies to the WLAN cards, some laptops have built-in cards, while other use Orinoco WLAN
cards plugged in the PCMCIA-slot. Each laptop also has a
GPS receiver connected via the serial RS-232 line. The GPS
devices are not only used for providing positioning information to the nodes, but we also use GPS for timing information. This information is provided once per second. The
GPS timing information is actually not required for the BLR
protocol, but only for performance measurements. The accuracy of the information is below 5 m and 200 ns for the
positioning and timing information, respectively.
In this paper, we present the results from experiments
that were conducted in the laboratory in order to validate
the implementation and for reference purposes, which allow a comparison with future outdoor experimental results.
As the GPS receivers do not work in indoor environments,
the position of the laptops had to be hardcoded to yield a
virtual topology. Therefore, the positions and the distances
between nodes in this virtual topology do not match the actual physical location of the laptops. Furthermore, all laptops are placed on a table within a few meters of each other
and thus could physically receive all transmissions of all
nodes. To ensure that a laptop only processes the packets
from laptops within the transmission range in the virtual
topology, a filter based on MAC addresses has been implemented. This filter operates directly on the pf_packet
socket and simply drops packets from out-of-range nodes
in order to match the physical and the virtual topology such
that the BLR application never sees packets from virtually
out of range nodes. This approach saves processing work
on the side of the BLR application since the kernel does
all the necessary filtering. The implementation of the MAC
filter is done by means of the Berkely Packet Filter (BPF)
language [11]. The GNU/Linux implementation is called
Linux Socket Filter (LSF) and is compatible with the BPF
Figure 4. Topologies for the experiments
Traffic is sent by the ping utility, which yields the round
trip time RTT. For each measurement, 2000 ICMP echo requests were sent, which together with the echo replies result in 4000 total data packets. The transmission rate had
to be limited to 10 echo request per second, because all
nodes are physically within each other transmission range.
A transmission of a node blocks all other nodes on the
MAC layer, and not only the neighbors in the virtual topology. Therefore, experiments with higher data rates where
a new IMCP echo request is sent out before the previous
echo reply arrived back at the source do not make sense.
The default packet size was set to 56 bytes. Including the
ICMP, IP, BLR, and MAC header this yields 180 transmitted bytes. The experiments were conducted with a rather
long M ax Delay of 5 ms and 25 ms to reduce the risk
that nodes transmit simultaneously due to the low interrupt
granularity as explained before in Section 4.5. The transmission range for calculating the Add Delay was set to a
250 m. Except for one experiment, we did not use the unicast mode in order to route packets as often as possible in
greedy mode. We used four topologies for the laboratory
experiments as depicted in Fig. 4, called chain, pairs, contention, and backup topology. Additionally, we also compared the measurements of these experiments with results
obtained from simulations conducted with the Qualnet [12]
network simulator and from analytical prediction. The scenarios for the simulations were identical to the experiments,
specifically the network topologies and the parameters of
the BLR protocol such as M ax Delay.
5.2 Chain Topology
The chain topology is the most simple topology as no
contention occurs and only one node always will forward
the packet. The forwarding node is located at the boundary
of the transmission range and almost immediately forwards
the packet without introducing Add Delay. Thus, the RTT
is basically independent of the M ax Delay as shown in
the histogram in Fig. 5, which shows the distribution of the
measured RTTs. The average is in both cases 17.4 ms and
the delivery ratio was always 100%. Considering the fact
that a packet pair is transmitted over eight hops (four hops
from the source to the destination for the echo request and
four hops back to the source for the echo reply), the measured RTT is approximately only 2 ms per hop. When we
roughly estimate that 180 Bytes are transmitted over 8 hops
with a bandwidth of 2 Mbps, we would expect an RTT of
approximately 6 ms. The Add Delay does only contribute
marginally to the RTT and is much less than 1 ms per hop,
because the progress of 249 m almost equals the transmission radius. In the simulations, we measured an RTT of
approximately 8 ms, which is close to the analytical estimation, considering that we did not take into account the
influence of the MAC layer. However, the RTT is only
about half the RTT measured in the experiments. The reason is that the Qualnet network simulator does not introduce any delay for processing packets at the nodes, i.e., the
Number of packets
Number of packets
mean = 29.618
Number of packets
mean = 83.684
The results from the pairs topology are given in Fig. 6
and Fig. 7. In this topology, a packet pair is again transmitted over eight hops. However, the RTTs now vary strongly
for the two different M ax Delay as expected. The two
transmissions from the nodes with only 20 m progress are
delayed significantly as they calculate a long Add Delay,
which is close to M ax Delay according to (1). In the
pairs topology, this yields RTTs of approximately 30 ms
and 80 ms for M ax Delay = 5 ms and M ax Delay =
25 ms, respectively. The delay introduced by BLR is approximately three times 230
250 · M ax Delay, two times from
S to D and only once from S back to D, which is approximately 14 ms and 69 ms for a M ax Delay of 5 ms and
25 ms, respectively. Together with the transmission delay
of 6 ms, we obtain an expected RTT of 20 ms and 75 ms
for the two different M ax Delay values. The respective
measured RTTs were 21 ms and 77 ms in the simulations.
These results are again approximately 8 ms shorter than in
the experiments, independent of the M ax Delay, which
confirms the previously stated reasons for the longer delay,
namely the zero processing time at the nodes in the simulator and the interrupt granularity of the Linux implementation.
In the pairs topology, we also evaluated the impact of
the unicast mode. Although packet duplication is not an
issue as only one potential forwarder exists, the RTT is af-
Max Delay = 5 ms
5.3 Pairs Topology
packets can be forwarded immediately, which is definitely
not the case in reality. Furthermore as mentioned in Section 4.5, the Add Delay has only a granularity of 1 ms in
the experiments. This is unlike for the simulations were the
Add Delay is really the calculated value and not ”rounded
up” to the next millisecond.
Delay [ms]
Delay [ms]
Max Delay = 5 ms and Max Delay = 25 ms
Delay [ms]
Max Delay = 25 ms
fected. Recall that in unicast mode, the packets are forwarded without introducing Add Delay if the next hop is
known. In Fig. 8, the histogram of the measured RTTs with
M ax Delay = 5 ms is shown. The RTT is significantly
shorter than when packets are always broadcast in greedy
mode and is reduced from 29 ms to 16 ms. We can also see
that there are some packet with longer RTTs around 25 ms.
The reason is that the unicast mode switches to greedy mode
every 5 s in order to detect possibly better located neighbors. Packets transmitted in greedy mode are again dynamically delayed at each node and not immediately forwarded
as in unicast mode.
5.4 Contention Topology
In the contention topology, three nodes receive the
transmitted packet from the source node and schedule the
packet for forwarding as they are all within the forward-
Number of packets
5.5 Backup Topology
5ms, with unicast
mean = 16.205
Delay [ms]
Figure 8. Pairs topology with unicast mode
and Max Delay = 5 ms
ing area. They calculate different Add Delay however and
the first transmitting node suppresses the others accordingly. In Fig. 9, the distribution of the RTTs is shown for a
M ax Delay of 5 ms. The results were almost identical for
M ax Delay = 25 ms due to the same reasons as for the
chain topology. We can observe that the RTT is quite short
compared to the previous investigated topologies because a
packet pair is only transmitted over four hops (two hops to
the destination and two hops back to the source). The the-
In a last experiment, we validated the backup mode of
BLR. There is no node located in the forwarding area and
the packets are routed in backup mode for three hops until arriving at the node closer to the destination than the
source. We measured two different RTTs of approximately
40 ms and 60 ms as depicted in Fig. 10. The reason is
that while the backup mode acquires neighbor information if greedy forwarding failed, i.e., during the beacon request reply dialog, other arriving packets are queued in the
backupqueue. When the backup mode setup is completed and the forwarding node has determined the next
hop by the “right-hand”, all queued packets are sent immediately to this next hop, thus, some packets in the queue
encounter shorter RTTs. The backup mode is also stateless
and does not store positions of neighboring nodes, therefore
the first following packet after the queue has been emptied
again has to wait until the request reply dialog is completed
in order to acquire the positions of the neighboring nodes.
The RTT is still quite short considering the fact that nodes
have to transmit a request for beacon packet and wait until
neighbors have replied. In the simulations, we measured an
RTT of 51 ms which is again approximately 1 ms less delay per hop than in the experiments due to the same reasons
as mentioned before.
mean = 50.807
Number of packets
Number of packets
mean = 7.849
.0 .9
64 -62
.0 .9
62 -60
.0 .9
60 -58
.0 .9
58 -56
.0 .9
56 -54
.0 .9
54 -52
.0 .9
52 -50
.0 .9
50 -48
.0 .9
48 -46
.0 .9
46 -44
.0 .9
44 -42
.0 .9
42 -40
Delay [ms]
Delay [ms]
Max Delay = 5 ms
Figure 9. Contention
Max Delay = 5 ms
5.6 General Observations
oretical transmission delay is approximately 3 ms because
of the reduced hop count. Since the Add Delay is again
for all nodes significantly below 1 ms, the expected RTT is
around 3 ms, which matches the measured RTT of 4 ms in
the simulations. The difference to the RTT of the experiments is again because of the required processing time at
the laptops.
Until now, we only considered the measured RTTs in
the experiments. We can conclude that the RTTs are short,
are as expected, and vary only slightly around the mean.
Other quantitative performance measurements are briefly
discussed in the following. In all four topologies, the delivery ratio was always 100%. This is not surprising consider-
ing the fact that the nodes are physically close to each other,
even if they are distant in the virtual topology. Furthermore,
we observed that in the chain, pairs, and contention topology, there was approximately one packet per thousand packets transmitted unexpectedly in backup mode. The reason
was that very rarely some packets showed a higher RTT and
collided with subsequent transmitted packets which caused
the required retransmissions in backup mode. This effect
is especially obvious in the laboratory where all nodes are
within transmission range. Furthermore, we observed very
infrequently duplicated packets, again in the order of some
few per thousand. However, they could be successfully suppressed at the next node by transmitting a control packet as
described in Section 4.2. Thus, no duplicated packet arrived at the destination node. Especially for the contention
topology, the few duplicated packets indicate that the first
transmitting node is able to successfully suppress the other
potential forwarders.
6 Conclusions
In this paper, we presented an implementation of the
position-based routing BLR on a GNU/Linux platform using laptops equipped with 802.11b WLAN cards and GPS
receivers. The advantages of BLR are that it is stateless
and does not require to have knowledge about its neighbors,
which allows the disposal of the periodical transmission of
hello messages and makes it immune to frequently changing network topologies. The BLR was implemented in the
user space of GNU/Linux and is transparently integrated
in the protocol stack, which allows to run arbitrary applications without modification. We discussed several problems encountered during the implementation and the experiments and described possible ways to solve them. BLR is
implemented to retrieve position information provided by
GPS receivers. Unfortunately, this information could not
be used in our laboratory experiments as all laptops were
within a single transmission range and a virtual topology
had to be configured manually. We conducted several laboratory experiments to validate the implementation. The forwarding of the packets in the greedy, unicast, and backup
mode of BLR was as expected. The results also indicate
that BLR is able to deliver packet over multiple hops in a
short time. Packets are forwarded reliably and the delivery
ratio was always 100%. Furthermore, the forwarding nodes
successfully suppressed the other potential forwarders and
acknowledge also the previous node reliably, because basically no duplicated packets were observed. In a next step,
we will conduct outdoor experiments and use GPS position information. In these experiments, the results may differ from the laboratory experiments because nodes may no
longer be within transmission range, which can cause duplicated packets and longer RTTs. We also plan to conduct
experiments with high mobility, for which BLR was originally designed, where the laptops are transported in cars.
[1] J. L. et al., “A scalable location service for geographic
ad-hoc routing,” in Proc. of MOBICOM ’00, Boston,
USA, Aug. 2000, pp. 120–130.
[2] P. Bose, P. Morin, I. Stojmenovic, and J. Urrutia,
“Routing with guaranteed delivery in ad hoc wireless
networks,” in Proc. of DIALM ’99, Seattle, USA, Aug.
1999, pp. 48 – 55.
[3] B. Karp and H. T. Kung, “GPSR: Greedy perimeter
stateless routing for wireless networks,” in Proc. of
MOBICOM ’00, Boston, USA, Aug. 2000, pp. 243–
[4] F. Kuhn, R. Wattenhofer, and A. Zollinger, “Worstcase optimal and average-case efficient geometric adhoc routing,” in Proc. of MobiHoc ’03, Annapolis,
Maryland, USA, June 2003, pp. 267 – 278.
[5] M. Mauve, J. Widmer, and H. Hartenstein, “A survey
on position-based routing in mobile ad-hoc networks,”
IEEE Network, vol. 15, no. 6, pp. 30–39, Nov. 2001.
[6] M. Heissenbüttel, T. Braun, T. Bernoulli, and
M. Wälchli, “BLR: Beacon-less routing algorithm for
mobile ad-hoc networks,” Elsevier’s Computer Communications Journal (Special Issue), vol. 27, no. 11,
pp. 1076–1086, July 2004.
[7] (2005, Apr.) Gentoo linux website. Gentoo Foundation, Inc. [Online]. Available:
[8] (2005,
Apr.) Virtual point-to-point(tun) devices. Maxim Krasnyansky. [Online]. Available:
[9] (2005, Apr.) The netfilter website. Harald Welte.
[Online]. Available:
[10] (2002,
ics Association (NMEA). [Online].
[11] S. McCanne and V. Jacobson, “The BSD packet filter:
a new architecture for user-level packet capture,” in
Proceedings of the 1993 winter USENIX conference,
San Diego, CA, USA, Jan. 1993, pp. 259–269.
[12] (2004, Nov.) Qualnet. Scalable Network Technologies
(SNT). [Online]. Available:
Implementation Strategies for a Secure and Efficient Multi-hop MANET
Minmin Tu, Jingyu Zhou, Guozhi Xu
Department of Electronics Engineering
Shanghai Jiaotong University
Huashan Road 1954
Shanghai 200030, P.R.C
{tuminmin,jyzhou,[email protected]}
Mobile Ad-hoc Network (MANET) is a multi-hop
wireless network without a preinstalled infrastructure,
which is attractive for commercial use due to its nature
of low cost and fast deployment. Routing and security
are two major challenges in building a real
commercial MANET environment. In this paper, we
focus on the software solutions to implement a secure
and efficient multi-hop MANET platform. We discuss
an efficient and stable model based on on-demand
routing protocols that is designed and implemented on
Windows CE with special considerations for security.
1. Introduction
Mobile Ad Hoc Network (MANET) [1] [2] is a
multi-hop wireless network without a preinstalled
infrastructure. Mobile nodes that are within each
other’s radio range communicate directly via wireless
links, while those that are far apart rely on other nodes
to relay messages as routers. Such network first arose
in disaster rescue and battlefield operations due to its
flexibility and relatively low cost. Nowadays, with the
prosperity of consumer electronic devices, such as
PDAs, mobile phones, set-top boxes, embedded
devices in cars, there is a trend to adopt ad hoc
networks for commercial uses.
Windows CE [8] is a commercial embedded
operating system designed for devices with small
amounts of memory, storage and CPU processing
power. It uses a component-based structure so that
Original Equipment Manufacturers (OEMs) can
choose only the operating system features that they
require for their specific hardware platform. Windows
CE also supports many multimedia applications and
games. In recent years many efforts have been put on
implementing MANET on Windows CE.
Routing model and security problem are two major
problems when applying ad hoc networks. The
infrastructure-less and dynamic nature of MANET
demands a new set of routing protocols. Existing
protocols to solve this problem include, but are not
limited to AODV [9], DSR [12], DSDV [13] and
TORA [14]. Most of the existing studies of these
protocols are only discussed in theory. Most
implementations to date are for Linux, and mainly for
experimental purposes. Reports on implementations in
Windows CE.NET for commercial use are much less.
On the other hand, security [3] in MANET is a hot
topic because the salient features of MANET pose new
challenges in securing the communication. Canned
security solutions like IPSec [15] [21] are not
In this paper we present the solutions to abovementioned issues from the point of view of the systemlevel software implementation. Focus has been put on
the software solutions to facilitate a secure and
efficient multi-hop MANET platform. We hope our
work will contribute to the proliferation of MANET in
real commercial world.
2. Existing implementations and related
There have been several implementations of ad-hoc
routing protocols. In this section, we provide a
comparison on these implementations.
V. Kawadia [4] studied the common requirements
imposed by MENET and proposed a general
framework with a generic API to facilitate MANET
protocol implementations and deployments. It is
known implementing the API normally require kernel
modifications. As a variation, they first provided an
implementation for Linux as a shared user-space
library called ASL(the ad-hoc support library) using
the standard Linux 2.4 kernel facilities and then
developed a full-fledged implementation of the AODV
protocol, called AODV-UIUC[7] using ASL.
Although this project was on Linux platform, the
major principles are still suggestive and helpful to our
design on Windows CE.
OLSR for Windows 2000 and Pocket PC is based
on the PICA library [16] [17]. The PICA (protocol
implementation specific library) serves to ease the
work of transplant of the Ad hoc Protocol on Linux to
the Windows 2000 as well as Windows CE platforms.
One flaw of the PICA is that it may introduce overhead.
UoBWinAODV [6] is a project for Windows XP.
It includes an NDIS intermediate driver and a user
space program to manage the operations of AODV.
However, UoBWinAODV meet with the problem of
low efficiency. For example, each data packet to be
sent out must search the matched record respectively in
two similar tables, one in the IP layer and the other in
the NDIS intermediate layer. In our model, we avoid
such repeated search in a long list by some
improvements in the NDIS driver.
WinAODV by Intel is similar to UoBWinAODV
generally and differs in the way that it can be used
together with ‘PUDL’, which assists WinAODV to
choose the best quality neighbor links.
Self-Organizing Neighborhood Wireless Mesh
Networks [5] for Windows XP is a community-based
multi-hop wireless networks implemented by
Microsoft networking research group, with features
including self-stabilizing, multi-path multi-hop
routing, auto-configuration and link quality
measurement, etc.
The implementations mentioned above are all
meant to address the on-demand routing problems.
Some of them, such as ASL and PICA, provide a
generic interface. Others are implemented for specific
operating systems. Here we only list the projects for
Windows, while those for Linux were enumerated and
analyzed by V. Kawadia in [4].
We also note that there were some projects
reported that support QoS like link quality
measurement and adaptive link choice, but few are
aimed at embedded operating systems like Windows
CE with consideration of security issue in practice as
we present in this paper.
3. Routing model
3.1. Routing protocols
In MANET, each node is capable of acting as a
router to provide end to end communication through
multi-hops. Most ad-hoc routing protocols can be
classified into two categories: proactive routing and
reactive routing. Proactive (or table-driven) routing
protocols are similar to those in the wired world, which
maintain routes to all possible destinations by
periodically exchanging control messages. Reactive
(or on-demand) protocols are designed specifically for
ad-hoc networks, which discover routes only when
there is such a demand for it. Obviously the second
mechanism requires less control messages.
Ad Hoc On-Demand Vector Routing (AODV)
protocol [9] is a typical on-demand routing protocol
for ad hoc mobile networks. The routing messages do
not contain information about the whole route path, but
only about the source and destination. Therefore, the
size of routed messages does not grow on the routing
path. It utilizes destination sequence numbers to
specify how fresh a route is and avoid loop freedom.
Experiments have shown that AODV performs
satisfactorily in most cases.
In AODV, a node updates its Neighbor List by
HELLO message, which is periodically broadcasted
from every node with lifespan equals to 1. Whenever
a node needs to send a packet to a destination for
which it has no fresh enough route, it broadcasts a
route request (RREQ) message to its neighbors. Each
node that receives the broadcast sets up a reverse route
towards the originator of the RREQ. When the
intended destination receives the RREQ, it replies by
sending a Route Reply (RREP). The RREP is
unicasted back to the originator of the RREQ. At each
intermediate node, a route to the destination is set.
Optionally, a Route Reply Acknowledgment (RREPACK) message may be sent by the originator of the
RREQ to acknowledge the receipt of the RREP. In
addition to these routing messages, Route Error
(RERR) message are used to notify the other nodes
that certain nodes are not anymore reachable due to a
link breakage.
3.2. Available solutions for Windows CE.NET
We conclude from the above projects, which most
on-demand routing protocols (like AODV) require the
following services:
a) Identify the need for a route request.
b) Notify ad-hoc routing daemon of a route request.
c) Queue outstanding packets waiting for route
d) Re-inject outstanding packets after successful
route discovery.
e) Update the routing cache.
However, most of current operating systems
including Windows CE.NET do not provide such
Table 1. Comparison of four methods
Port AODV using
Based on the
modified kernel
driver and a user
routing program
All in the NDIS
Easy to
A long solution
A new generation
of OS
Efficient and
Easily turn on/off
by users
Efficient and
Time delay
Infeasible now
Relative codes are
not open.
Difficult to
control by users
Based on the above analysis, we set out our
implementation based on the routing model using the
third method.
3.3. Routing model
Figure 1 illustrates the structure of this
implementation. Two main components are the
AODV routing daemon and the NDIS IM driver, both
of which run in the user space.
implemented as an executable (.exe) file so that the
users can choose to run it only when needed. NDIS
IM driver is a dynamic link library (.dll) file, which is
loaded and bound to the specific net card according to
the registry when the operating system starts up.
These two components share the same memory space
for outstanding packets (data packets without any
valid route), control packets and route usage
information. They synchronize with each other using a
pair of sophomore. AODV synchronizes its internal
route cache with the IP route table by reading and
updating the IP route table using the IP Helper tools.
AODV sends out packets using WinSock DLLs.
Send control packets
Resend data packets
Win Sock
route table
NDIS Protocol Driver
NDIS IM Driver(.dll)
NDIS Miniport Driver
IP Layer
services explicitly. To address these problems, we first
propose four major methods for comparison.
The first method is to port the existing AODV
implementation for Linux to Windows CE.NET using
PICA mentioned in Section 2. It is an easy and
straightforward, but overhead and time delay will be
The second method is to change the kernel of
Windows CE.NET so that it can support the MANET
protocols directly. It is a long-term fundamental
solution. However, since the codes for IP layer in
Windows CE.NET are not open to the public, we do
not apply this idea in this stage.
The third method is to adopt an NDIS intermediate
(IM) driver and a user routing program. The Network
Device Interface Specification (NDIS) is a public
interface, which governs the communication between
interface device drivers controlling hardware adapters.
The NDIS intermediate driver is located between the
NDIS miniport drivers and NDIS protocol drivers for
some extended functions such as filter or
encryption/decryption (Figure 1). The user routing
program manages the routing environment as required
by the AODV protocol. It is thus up to the user to
decide when to turn on or turn off the routing program.
The fourth method is to implement all the
functions in the NDIS intermediate driver. This
method is feasible because Windows CE removes the
barrier between kernel space and user space for device
As all the networking device drivers
effectively run in the protected user mode, they can be
linked with the WinSock DLLs directly and send the
AODV control packets as UDP datagram from port
654 to other host using WinSock functions. One
drawback of this method is that it will increase
complexity for users to control the routing model
because it is in essential a driver rather than an
executable file.
Capture packets
Trigger route discovery
Update expired time
Figure 1. Software architecture of routing
We find that the Win Sock DLLs in Windows
CE.NET do not support ‘raw’ socket. This adds to the
complexity for implementing AODV when it re-sends
the outstanding packets, because the transport layer
type and port of those packets vary over different
applications. Figure 2 illustrates our solution to this
Ctrl Msg
Win Sock
TCP/IP Stack
NDIS IM Driver
Figure 2. Outstanding Packet Flow
UDP port 654 is the reserved port for AODV. To
be compliant, we define UDP port 655 for the reinjected data packets. Therefore, NDIS IM driver
could distinguish the re-injected data packets from the
control packets by their UDP port number. As shown
in Figure 2, queued outstanding packets in AODV
include the source IP header, the TCP/UDP header
and data. After successful route discovery, we reinject the outstanding packets and send them through
UDP port 655 without consideration of the source
TCP/UDP type or port number. Hence, all parts of the
source outstanding packets are regarded as payload by
the TCP/IP stack. New UDP header (UDP0) and IP
header (IP0) are added as well as the MAC header.
While in the NDIS IM Driver, the redundant headers
IP0 and UDP0 are cut and re-packed. Although the
re-pack operation in the NDIS IM Driver is a little
time-consuming, it will not impact the whole
performance because this only affects re-injected
outstanding packets.
of searching and comparing with a single comparison.
We improve the searching effectiveness by setting a
static ARP record in the system and comparing the
MAC address rather than IP address in the NDIS IM
In the original version, route daemon updates not
only the IP route table but also the send filter list in
NDIS intermediate driver. When a data packet passes
the NDIS intermediate layer, its destination IP address
is compared with those in the send filter list. If there is
no match, then this data packet is regarded as an
outstanding packet and should be sent to routing
In the improved version, we assume that all the
mobile nodes in the MANET have the same virtual
gateway, which does not represent a concrete device.
Hence, the next-hop of default route entry in the IP
route table is the IP address of the virtual gateway.
We set a new ARP entry. The default gateway
address maps a pre-defined MAC address VMAC
Ip Addr:
MAC Addr:
Thus, when a data packet passes the IP layer
without a route match, it is then by default set to the IP
address entry, specifying that the destination
MAC address of any outstanding data packet is VMAC.
So, in the NDIS IM driver, the data packet only need
to compare its destination MAC address with VMAC
rather than the IP address filter list that is much longer.
The second improvement is to choose a more
efficient search algorithm in the AODV routing
daemon. In a flat MANET with no sub-net, the
conventional linear search of hash tables is suitable
here because there is no need for prefix matching.
4. Security solutions
3.4. Improve the search efficiency
4.1. Security issues in MANET
All the improvements described below can be
applicable in a simple, flat and trusted MANET
environment without sub-net or compromised nodes.
We find out that the bottleneck of the model lies in
the search algorithm, which is used both in the send
filter function of NDIS IM driver to filter outstanding
packets as each packet passes down and in the AODV
routing daemon when processing the route control
packets and outstanding packets. We reduce the
search cost in two ways.
Since the search in the NDIS IM driver occurs for
each packet, a better solution is to replace the process
As Zhou indicates in [3], there is an increasing
possibility of eaves-dropping, spoofing, and denial-ofservice attacks on MANETs, due in part to their poor
physical protection and dynamic topology and
dynamic membership. Some security-enhanced
versions of ad-hoc routing protocols, such as
SAODV[10], SRP[18], ARAN[19] and SEAD[20],
have been proposed to ensure the operation of the
routing protocol unaffected by attempts to forge or
alter the routing protocol control messages.
Meanwhile, data packets could be protected by
conventional encryption methods.
However, Public Key Infrastructure (PKI), or
more basic key exchange techniques are difficult to
implement in an ad hoc network due to the lack of
authorities of trust and appropriate network
infrastructure [11].
Many schemes have been
proposed, yet a full-fledged and feasible key
management solution is still underway. In our work,
we suggest and demonstrate that the key distribution
approaches could be tailored for Ad Hoc network to
different scenarios.
4.2. Security services provided by Windows
Windows CE supports Secure Sockets Layer (SSL)
security protocols for secure network communication.
SSL supports are available directly from WinSock.
These applications use secure sockets to send and
receive encoded data over the communication lines.
IPSec is designed to encrypt data as it travels
between two computers, protecting the data from
modification and interpretation. It is a key line of
defense against internal, private network, and external
attacks. Windows CE-based IPSec does not provide
programming interfaces for developers.
Cryptography is employed to ensure data integrity
and secures communication in the Windows CE-based
applications. The Microsoft cryptographic system
consists of application, operating system (OS) and
cryptographic service provider (CSP). Applications
communicate with the OS through the cryptographic
API (CryptoAPI). The OS communicates with CSPs
through the cryptographic service provider interface
WindowsCE.NET also provides security services
for user authentication, credential management, and
message protection through a programming interface
called the Security Support Provider Interface (SSPI).
Within SSPI, different security providers are available,
such as the NTLM security support provider (SSP) and
Kerberos SSP.
Each one contains different
authentication and cryptographic schemes.
4.3. A secure routing model
In this section, we integrate our security solutions
into the routing model described in Section 3. It is
assumed that ad hoc nodes have permanent addresses
and are in a flat MANET without sub-nets (E.g., the
net meeting scenario). Therefore, bindings between
public keys could be achieved by free distribution or
by a certification authority (CA) to issue public key
certificates. Figure 3 illustrates the structure of our
secure routing model.
SAODV (secure AODV) is a secure version of
basic AODV routing protocol, which achieves those
security goals: import authorization, source
authentication and integrity. Two mechanisms are
therefore used: digital signatures to authenticate the
non-mutable fields of the message, and hash chains to
secure the hop count information (mutable field of the
message). In Windows CE.NET, Crypto APIs are
used to provide the required digital signature and hash
chain functions.
IPSec is used to secure the network data
transmissions in MANET. It is necessary that the
IPSec policy will be able to apply certain security
mechanisms to the data packets and just bypass the
routing packets, which typically can be identified
because they use a reserved UDP port number.
However, Windows CE-based IPSec does not provide
programming interfaces for developers to perform such
specific operations.
Hence, the Windows CE-based IPSec is disabled
and some IPSec functions are embedded in the NDIS
IM driver. Detailed operations are described with the
pseudo-code (see Figure 4 and Figure 5).
Furthermore, some sensitive applications may
choose the secure socket (support SSL) to transmit
Hash &
(support SSL)
Win Sock
IP Layer
Figure 3. Structure of a secure routing model
if (UDP & UPD port == 654)
then {
send to AODV} //route control message
else { decrypt,
pass up} //data packet
Figure 4. Pseudo-code of NDIS IM receive part
if (MAC == Default_MAC)
then { block,
send to AODV} //packet with no route
else if (UDP & UDP port == 654)
then { pass down } //route control message
else if (UDP & UDP port = 655)
then { cut IP0 and UDP0 from the packet,
encrypt, //if use IPSec
record the route used, //inform AODV
pass down} //resend data packet with
//valid route
else { encrypt, //if use IPSec
record the route used, //inform AODV
pass down} //other packets with valid
Figure 5. Pseudo-code of NDIS IM send part
5. Conclusion and future work
In this note, we study several alternative
implementations of on-demand routing protocols in
Windows CE.NET platform, and propose an efficient
routing model in a specific embedded operation system.
Then, we explore the security problems for MANET,
and integrate our security solutions into the routing
model with the help of the security services provided
by Windows CE.NET.
We believe that routing and security are two major
issues in building a real commercial MANET
environment. In this paper we present the solutions to
those issues from the point of view of the system-level
software implementation. With our secure routing
model embedded, several mobile nodes could form a
secure and efficient multi-hop MANET, in which some
application scenarios like net-meeting could be
realized. We hope our work will contribute to the
proliferation of MANET in real commercial world.
In the future, we plan to extend our work to more
complicated yet more practical schemes, such as
dynamic address allocation, auto configuration and etc.
The corresponding key management solutions, which
are laid aside for the moment, will be put into further
6. References
[1] K. Kuladinithi and C. Gg, “Tutotrial on mobile ad hoc
networks”, in Proc. of 1st Regional Conference on ICT and
E-paradigms, June 2004.
[2] Manet working group charter. [Online]. Available:
[3] L Zhou, Z J Haas. “Securing Ad Hoc networks”, in IEEE
Network Magazine, 1999,13(6):24~30
[4] V. Kawadia, Y. Zhang, and B. Gupta., “System services
for ad-hoc routing: Architecture, implementation and
experiences”, in Proc. of the 1st Int’l Conf. on Mobile
Systems, Applications, and Services (MOBISYS), May 2003.
[5] Self-organizing neighborhood wireless mesh networks.
[Online]. Available:
[6] Uobwinaodv implementation. [Online]. Available:
[7] Aodv-uiuc implementation. [Online]. Available:
[8] MSDN. [Online]. Available:
[9]C. E. Perkins, E. M. Royer, and S. R. Das. “Ad hoc ond e m a n d d i s t a n c e v e c t o r ( A O D V ) r o u t i n g ” . IETF
INTERNET DRAFT, MANET working group, Jan. 2002.
[10]M. Guerrero. “Secure ad hoc on-demand distance vector
(SAODV) routing”. IETF MANET Mailing List,
[email protected], , Oct. 2001.
[11]N. Asokan and P. Ginzboorg. “Key agreement in ad-hoc
networks”. Computer Communication Review, 23(17):1627–
1637, Nov. 2000.
[12]D. A. Maltz. On-Demand Routing in Multi-hop Wireless
Mobile Ad Hoc Networks. PhD Thesis, Carnegie Mellon
University, 2001.
[13]C. E. Perkins and P. Bhagwat. “Highly dynamic
destination-sequenced distance-vector routing (DSDV) for
mobile computers”. In Proceedings of ACM SIGCOMM’94,
London, U.K., Sept. 1994.
[14]V. D. Park and M. S. Corson. “A highly adaptive
distributed routing algorithm for mobile wireless networks”.
In Proceedings of IEEE INFOCOM, 1997.
[15] S. Kent and R. Atkinson, “IP Authentication Header,”
IETF RFC 2402,1998.
[16] C. M. T. Calafate and P. Manzoni, “A multi-platform
programming interface for protocol development”. Eleventh
Euromicro Conference on Parallel, Distributed and
Network-Based Processing, 2003, Genova, Italy
[17] C. M. T. Calafate, R. G. Garcia and P. Manzoni,
“Optimizing the implementation of a MANET routing
protocol in a heterogeneous environment”. Eighth IEEE
Communications, 2003, Kemer-Antalya, Turkey
[18]P. Papadimitratos and Z. J. Haas, “Secure routing for
mobile ad hoc networks”. SCS Communication Networks
and Distributed Systems Modeling and Simulation
Conference (CNDS 2002), Jan 2002.
[19]B. Dahill, B. N. Levine, E. Royer, and C. Shields, “A
secure routing protocol for ad hoc networks”. Technical
Report UM-CS-2001-037, University of Massachusetts,
Departament of Computer Science, Aug. 2001.
[20]Y. C. Hu, D. Johnson, and A. Perrig, “SEAD: Secure
efficient distance vector routing for mobile wireless ad hoc
networks”. In Fourth IEEE Workshop on Mobile Computing
Systems and Applications (WMCSA ’02), June 2002, pages 313, June 2002.
[21] S. Kent and R. Atkinson, “IP Encapsulating Security
Payload (ESP),” IETF RFC 2406, 1998.
A linux based Bluetooth scatternet formation kit: from design to performance
Francesca Cuomo, Andrea Pugini
INFOCOM Dept., University of Roma “La Sapienza”, Via Eudossiana 18, 00184, Rome (Italy)
[email protected], [email protected]
Bluetooth is a key technology for the deployment of
wireless personal (WPAN) and body area (WBAN)
networks. This paper describes an implementation
experience of a scatternet formation protocol on a
Bluetooth stack for Linux (BlueZ).
The innovative elements of this work are: I) the
development of two software packages designed to
create and manage multi-hop Bluetooth wireless
networks, based on simple USB Bluetooth devices; II)
the analysis of results on this real implementation.
The formed network (scatternet) presents limited
formation delay and interesting performance in view of
personal/body area applications. Implementation
details as well performance results are discussed.
These results have been derived in the framework of
the FIRB VICOM project founded by the Italian
Telecommunication Ministry.
1. Introduction
Bluetooth is a short range, low power and low cost
wireless technology for ad-hoc networking. Eight
Bluetooth devices are interconnected in a piconet and
share the same radio channel which supports about 1
Mbit/s [1] nominal symbol rate. Class 3 devices have a
transmission range of about 10 meters. No network
infrastructure is envisaged: self-organization and peer
communications lead to a complete ad-hoc
connectivity. Piconets are dynamically set-up and torn
down. In a piconet two Bluetooth devices (named also
nodes) exchange information by means of a masterslave relationship. Master and slave roles are dynamic:
the device that starts the communication is the master,
the other one is the slave. A master can connect with
up to 7 slaves. Multiple access in a piconet is centrally
regulated by the master that adopts a polling scheme
on a slotted time structure. Piconets can coexist in the
same area and interconnect in a scatternet.
A Bluetooth device that joins more than one piconet is
named bridge and participates to communications in
different piconets on a time-division basis. A device
can play the role of master in only one piconet.
Scatternet formation in Bluetooth has recently received
a significant consideration [2]. Existing solutions can
be classified in single-hop ([3][4][5]) and multi-hop
In this paper, we report our experience using the
Bluetooth stack for Linux (BlueZ). We developed
innovative software applications with graphical user
interface, called J-BlueZ and BASkit. The former
translates visual inputs in BlueZ commands, the latter
supplies automata to dynamically form and manage
scatternets. Our study includes testing and
performance evaluation of these packages to collect
relevant results about implemented scatternet
formation protocol. These packages are used for
mobile immersive communications to allow a user,
managing simple devices (e.g., PDAs and smart tags),
to enter and act in pervasive environments.
2. The Bluetooth Profiles
The Bluetooth standard defines all the protocols
from Bluetooth Radio layer to upper network layers.
Also different profiles are specified [10][11].
The Personal Area Networking (PAN) profile
describes how two or more Bluetooth enabled devices
can form an ad-hoc network and how the same
mechanism can be used to access a remote network
through a network access point. The profile specifies
the roles: the Network Access Point (NAP), the Group
Ad-hoc Network (GN), and the Personal Area
Network User (PANU). NAP can be a traditional LAN
data access point while GN and PANU are master and
slave nodes in a piconet.
The PAN profile defines a means of enabling
Bluetooth devices to participate in a personal area
network. Completely un-modified Ethernet payloads
can be transmitted using the Bluetooth Network
Encapsulation Protocol (BNEP) to exchange packets
between Bluetooth devices (Figure 1).
• OBEX Object and File Push Profile.
Figure 1: BNEP profile
This profile defines how PAN is supported in the
following situations:
• Ad-hoc IP networking by two or more Bluetooth
devices in a single piconet;
• Network access for one or more Bluetooth devices.
BNEP specifications describe the protocol to be
used by the Bluetooth PAN profiles. They also define
a packet format for Bluetooth network encapsulation
used to transport common networking protocols over
the Bluetooth media. Bluetooth network encapsulation
supports the same networking protocols that are
supported by IEEE 802.3/Ethernet encapsulation.
Packets from the supported networking protocols are
contained in Bluetooth network encapsulation packets,
which are transported directly over the Bluetooth
L2CAP protocol.
3. BlueZ Description
Many implementations of the Bluetooth stack are
available, but not all of them are free or open source.
Open source implementations are free with easy
access to the API’s and source code. General Public
License makes sure that the software is always
designed in a open manner and this may guarantee
many open source developments in the future.
BlueZ is the Official Linux Bluetooth protocol stack
for the Linux Kernel [12].
Some of BlueZ’s features are (Figure 2):
• Support for core Bluetooth layers and protocols;
• Support for multiple Bluetooth devices;
• Standard socket interface to all layers;
• Multiplatform: x86 (single and multiprocessor),
SPARC, ARM, DragonBall;
• Support for L2CAP, SCO, RFCOMM and SDP;
• Support of GAP, DUN, LAN, SPP, PAN profiles;
Figure 2: BlueZ stack
BlueZ is distributed in a set of packages . BlueZ
core package, called “bluez-kernel”, includes all
functionalities to set up the core of Bluetooth. It takes
care of making HCI-devices, the L2CAP and LMP/LC
protocols. It is already a part of 2.4.21 kernel or
higher. [12]. Other packages complete BlueZ: “bluezlibs”, “bluez-utils”, “bluez-sdp”, “bluez-pan”, “bluezhcidump”, “bluez-hciemu”, “bluez-bluefw”.
3.1. BlueZ tools
BlueZ contains three important tools to manage
Bluetooth devices.
The first BlueZ tool is hciconfig that makes possible
to configure device at the HCI level.
The hciconfig main call for the first Bluetooth
interface is ‘hciconfig hci0’; this acts like the ‘ifconfig’
network command in Linux. To set up a device, the
first call must be: ‘# hciconfig hci0 up’.
Another tool is hcitool. This tool is employed to use
the main features of device at HCI level (e.g., inquiry
and connecting calls).
Third tool is PAND, treated in the following.
3.2. PAND
BlueZ package implementing PAND contains the
necessary to use two of the most used profiles in
Bluetooth: LAN and PAN. The PAN profile specifies
the three Bluetooth roles:
Network Access Point (NAP): acts as proxy,
router or bridge between an existing network
infrastructure (typically LAN) and (up to 7 active)
wireless clients (PANUs);
• Group ad-hoc Network (GN) controller: entity
interconnecting up to 7 (active) PANUs (master in
the Bluetooth Piconet);
• PAN User (PANU): client of a NAP or client-type
member of a GN. It corresponds to the slave role
of the Bluetooth Specifications.
Both NAP and GN are service providers that
coordinate the traffic in the PAN. NAP is also an
access point to another network. The PANUs, on the
other hand, are clients or users of the PAN Service.
The Bluetooth PAN feature offers (among other
features) IP support over Bluetooth (i.e., L2CAP).
3.2.1. PAND: General Setup. We consider the
configuration of a single link. The master of the link is
configured as the GN of the link, the slave as the
PANU. The basic idea behind PAND is that you start it
on one side in a server fashion (configuring the master
side of the link), using the '--listen' parameter and then
you can establish an explicit connection between
PANU and GN.
At the end of the PAND configuration, each
interface activated between two nodes represents a
wireless link and is named ‘bnepx’ (with x=1,2,3..).
3.2.2. PAND interfaces. After the PAND
connection is established, a virtual network interface
'bnep0' is created on both nodes. This interface can be
configured using 'ifconfig' (e.g., ‘# ifconfig bnep0’) to associate the IP address of the node to the
interface. The handling of the ‘bnepx’ interfaces could
become complicate in case of multihop paths. In this
case, a node belonging to a path should receive from
an interface (e.g., ‘bnep2’) and forward on a different
interface (e.g., ‘bnep4’). This means that the node
needs a routing table and this table must be
dynamically changed when new connections are
established in the scatternet.
3.2.3. PAND Ethernet Bridging. In order to
simplify the aforementioned situation, it is possible to
use the "802.1d Ethernet Bridging" contained in the
2.4.x Linux kernel. The idea of this feature is to
associate the multiplicity of Bluetooth layer 2 network
interfaces (‘bnepx’) to a single layer 3 interface named
Each device associates its IP address to this
interface, and associates all the communications
related with the interface ‘pan0’ to the BNEP protocol.
This procedure has to be repeated at each node for
each interface ‘bnepx’ created by PAND.
Once the ‘pan0’ interface is created it is the unique
interface to link any 'bnepx' interface. The composition
of the different ‘pan0s’ in the nodes constituting the
piconet appears as a unique “piconet broadcast”
interface. Then, to use the scatternet as a unique
broadcast domain you need to associate the BNEP
broadcast protocol to the pan0.
4. Modules and Interfaces using BlueZ:
from J-BlueZ to BASkit
BlueZ supplies a large range of services and profiles
to create and manage Bluetooth ad-hoc connections
but requires to act single commands using shell
terminal. Such procedure is slow and complex: it
requires complete knowledge of BlueZ sintax and
semantic. To facilitate the use of BlueZ in a scatternet
test-bed framework, we developed a Java Graphical
User Interface (GUI) for BlueZ: J-BlueZ. This is an
intuitive, essential, graphic interface with basic
capabilities over BlueZ. It basically does not
implement any network formation protocol but gives
control to the user, merely transforming visual
commands in series of BlueZ commands. J-BlueZ
executes commands invoking shell terminal and
providing to it a string containing commands. As it
happens when the interface is executed by human user.
Beside the GUI, a second software package was
developed: BASkit (Bluetooth Ad-hoc Scatternet kit).
This package is a network module providing J-BlueZ
simplicity and effectiveness but implementing a
concrete Bluetooth ad-hoc network formation protocol.
The goal is twofold: create dynamic and optimized adhoc Bluetooth network, and manage and monitor
connections in this network. It supplies multi-hop
connectivity and supports TCP and UDP sockets, IP
broadcast and ICMP to upper layers.
Therefore, J-BlueZ and BASkit can be used to get
connectivity with a Bluetooth device and present other
relevant advantages for network managers and
deployers: J-BlueZ can be easily used being unaware
of components as operative system or BlueZ; BASkit
can be also used in a test-bed to verify implemented
Bluetooth ad-hoc network formation protocol and
relevant performance.
4.1. J-BlueZ
4.1.1. Usage. The J-BlueZ application module was
designed to be a simple and intuitive tool to create and
manage network connections. According to this
guiding principles, the interface provides the core
• collecting available devices information;
• connecting to/ disconnecting from a device;
• setting up parameters.
User can act using buttons and text fields creating
and tearing down connections by himself or select one,
or more, of “Automatic” features and let application
create and manage network connections as a wizard.
4.1.2. Structure. J-BlueZ structure is designed to
accomplish few essential goals:
• to display available devices information;
• to make possible connection and disconnection
to/from a remote device.
While working, the application uses some
parameters to lead its behavior, user can modify these
parameters and set some procedures as autonomous or
J-BlueZ has a unique main menu (Figure 3)
• device information area, showing available devices
• a series of buttons and text fields to input
Figure 3: J-BlueZ main window
In device information area, a row for each available
device shows: the address, the name, the connection
role and the BNEP interface (if device is connected).
Parameters menu is in the upper left corner of the
window (Figure 3): each menu item activates a dialog
window to change parameter value. “Time interval”
item can be used to input a value for the time interval
of the periodic environment inspection: typing ‘0’
(zero) periodic environment inspection is disabled. In
this dialog box the user can also activate environment
inspection after every command execution. Another
feature is “Automatic Behavior”: this radio button
enabled, the module acts as an autonomous tool, as
describe in the following.
4.1.3. Behavior. J-BlueZ is a general purpose
module, its main goal is to simplify the execution of
PAND commands and the inspection of network
information providing the user with basic options to do
so. Application does not implement any network
formation protocol by itself.
When application is launched it enables potential
incoming connection from other devices. This phase is
repeated periodically and continues during all the life
time of J-BlueZ. In every moment the user can decide
to connect to a device and become, in this way, a slave
of that device. Therefore, many connections are
possible at the same time with different roles for a
device. Commercial devices employed in our test-bed
support multiple L2CAP and multiple roles.
J-BlueZ does not add any additional control
features on connection: if too many connections are
attempted J-BlueZ will simply communicate an error.
In case of failure, short messages are displayed and
information about environment and network status are
not updated.
Information about available devices are provided
after environment inspection procedure and are
maintained in the devices text area until new
environment inspection is performed. Environment
inspection is performed:
• each time user clicks the “refresh info” button,
• each time periodic inspection time interval expires
(if it is greater than zero),
• each time a user command is executed and
“Automatic Inspection” is selected.
During inspection, as all other procedures,
application appears as “freezed”: buttons and text
fields are inactive and information do not change
anymore. Only when the procedure is accomplished
information are updated and buttons are reactivated.
In order to create connections an IP address must be
assigned to the Bluetooth device by the J-BlueZ
application. At the launch, in the startup dialog, the
application asks for this address and the device name.
As we mentioned before an automatic behavior is
possible. Activating this feature makes J-BlueZ a
completely autonomous connection wizard: “Connect”
and “Disconnect” buttons are disabled and application
automatically inspects environment, using defined time
interval, and creates connections to every available
device. In this way the host is eventually linked to each
device creating more links than necessary i.e.,
generating a redundant topology.
J-BlueZ simplicity and effectiveness is therefore
preferred in the tradeoff with network efficiency. This
network efficiency is delegated to the BASkit module
as described in the following.
4.2. BASkit
BASkit (Bluetooth Ad-hoc Scatternet kit) includes
two logical entities:
• network module for dynamic and optimized ad-hoc
Bluetooth network formation and management;
• Graphical User Interface for visual direct control
and action capability over entire network.
The GUI is not strictly linked to network module: the
latter must be first launched, or loaded in background
as service, and then GUI can be activated by a shell
command or by double clicking on its icon.
4.2.1. Network module. Network module creates
and manages the network, providing “plug-and-play”
The network module acts in two steps:
• it adds new nodes to the network using a fast and
efficient approach;
• it manages and organizes the network topology by
applying a suitable algorithm.
This algorithm, called SHAPER [8], ensures treelike network topology and it enables the user to add
new nodes to the network, merge different trees,
reorganize automatically the network topology after
nodes abandon event (so called self-healingness). To
do so it uses short messages to and from nodes, sent
over the network through specific sockets.
When a host activates BASkit (network module) it
attempts to connect to other devices, that is already
existing network, by:
• applying the J-BlueZ “inspection” procedure,
• connecting to one device among those discovered,
randomly chosen.
Such procedure can significantly reduce the time
needed to be connected to the network, avoiding the
long time required by multiple single link creation in
BlueZ. Such delay is caused by Bluetooth usage of a
single beacon channel and BlueZ specific Bluetooth
implementation and therefore cannot be changed
without modifying BlueZ code. The BASkit module
overcome this issue by reducing the number of
connection attempts without modifying any BlueZ
When a BASkit host (A) creates its first connection
it receives from the connected host (B) tree-like
structure information, according to SHAPER
algorithm. At this step different scenarios can occur:
• Host “B” belong to a tree-like structure - “A”
becomes a member of the already formed network
and follows SHAPER rules and procedures;
• Host “B” is isolated - Hosts “A” and “B” create a
single leaf tree-like structure and become a new
Once nodes are connected in a network, connectivity
given by BlueZ enables use of sockets to send and
receive messages required by SHAPER. BASkit
network module is a “scatternet” manager since
manage logical tree-like structure: it creates scatternet
using piconets.
4.2.2. Graphical User Interface. The BASkit GUI
can be activated by the user when he/she needs to have
control on module activity or be aware of the network
status (additional features to network connectivity).
This interface provides complete network description
and intervention capability, it is a key tool for network
deployer and manager to control and to act on entire
Bluetooth ad-hoc network from a single node.
The BASkit GUI development process followed
some fundamental guidelines:
• we need detailed information about network and
• we want to monitor module working, with its
sockets, channels and messages;
• we want to be able to directly act on the network.
These requirements defined the GUI layout: graphic
pane displaying visual network structure, text areas
showing information about network nodes and module
work, buttons and text fields to input commands. Some
elements are inherited from J-BlueZ and other realized
for the first time.
BASkit implements scatternet formation protocol in
a tree-like form and uses this structure in the GUI to
display scatternet nodes and node relative information
(figure 4).The resulting Panel, even if giving in one
shot many information, has a high intelligibility in any
information displayed.
Figure 4: BASkit panel showing scatternet
tree-like structure
Thanks to this simple and clear picture we identify
piconets within scatternet. The highest node (“Teacher
PC” in Figure 4) is the master of first piconet, it is the
scatternet root. Devices listed below are slaves (and
leaves). Devices master of other piconets are marked
with “-“ and have other devices listed below.
BASkit GUI, thanks to the text panel, provides
information about the local Bluetooth host device
features such as: type, address, status, traffic class,
packet loss, other information distributed among the
scatternet nodes.
Also BASkit interface has many J-BlueZ features,
even if empowered, like connection, disconnection and
network inspection, but procedures are implemented in
different ways through different BlueZ instruments.
Interface enables sending to remote node of
information request and specific commands forcing it
to execute operations. Manager, using such instrument,
can send over network commands to remote nodes
having complete control of network and nodes activity.
There is a boundary to this control: remote nodes
execute received commands only if their GUI is not
5. Experimental results
In this Section we evaluate:
• performance of the BlueZ stack with respect to link
throughput, as well as latency, jitter and packet
• performance of J-BlueZ and BASkit with respect to
time required to create network.
Here, the number of slots used for transmission and
the resistance to noise have a significant impact. Our
result shows that the best throughput is achieved with
DH5: 5 slots transmission and lower noise resistance
(more data is transmitted than other codings).
Figure 5: Mean throughput of a FTP over
Note that the measured maximum throughput is far
from the theoretical maximum (723 Kbit/sec). In fact,
our BlueZ stack generates messages at different layers
that take up bandwidth on the user interface.
5.1. BlueZ performance
We build up the experimental platform with actual
Bluetooth commercial devices. The test hardware is
composed by personal computers: two Pentium 4 CPU
(1,8Ghz, Mandrake Linux 9.1) and one Pentium Pro
CPU (800MHz, Linux 9.2). Commercial USB
Bluetooth dongles, 3Com and Nortek brands, were
All the Bluetooth dongles can be bridges also with
different roles, e.g. PANU/GN or PANU/PANU.
5.1.1. Piconet and Scatternet Performance
Evaluation. We measured throughput on point-topoint connection between a master and a slave in a
single piconet. Figure 5 shows the mean throughput for
all possible Bluetooth packet encodings. The
throughput evaluations have been carried out by
executing an FTP of a file of 10 Mbytes.
Figure 6: Topology of the scatternet used in
the experiments
Then we measured throughput on point-to-point
connection in different scenarios. In the first one we
set up a piconet with one master and two slaves. In the
second one, we have set up a scatternet connecting two
piconets formed by two units (Figure 6). The bridge
node plays different roles (GN – PANU) in the
piconets. In the third case, we have two piconets,
formed by master and slave, but bridge plays same role
(PANU) in the piconets.
FTP mean throughput is: about 550 kbit/sec over 1hop piconet, about 300 kbit/sec over 2-hops PANUGN/PANU-GN scatternet, about 175 kbit/sec over 2-
hops GN-PANU/PANU-GN scatternet and about 60
kbit/sec over 2-hops piconet.
Scatternet tests have different throughput and this is
caused by bridge node roles: a PANU-PANU bridge
node has the same throughput in piconets but a GNPANU bridge node asymmetrically allocates activity
and causes a throughput reduction.
while a coding for an audio communication can be set
up to 10 kbit/sec.
To evaluate the performance of the wireless network
with Bluetooth, we have used the Gnomemeeting
video-conferencing application over a two-hops
scatternet. The video coding used had as maximum
bandwidth 100 kbit/sec, while the audio coding is up
to 10 kbit/sec. Gnomemeeting uses UDP packets. To
have a term of comparison we also ran an FTP for a 10
Mbyte. We measured the packet loss and the jitter.
Figure 7: Mean throughput of a FTP over a
two-hops piconet and scatternet
This experiment can primarily suggest to use, with
BlueZ, multi hop scatternet topology rather than a
single piconet. This result encouraged the design of
suitable tool for the automatic scatternet formation
5.1.2. Audio-Visual Evaluation. A current
challenge for video over a wireless link, using devices
such as Bluetooth, is the high degree of variability in
the radio signal strength meaning unreliable
connections between devices. When sending video
using Bluetooth connections, these unreliable
connections can cause errors which result in huge
defects in the quality of the image received. As for a
two-hops connection, since the nominal mean
throughput for a two-hops scatternet connection is 150
kbit/sec, it should represent a critic value. On the
contrary, in case of two-hops connection over a
piconet, the bit rate (60 kbit/sec) is not sufficient to
support a videoconferencing application. From these
considerations, we can notice that on a piconet built by
a single link it should be possible to have a videoaudio connection over Bluetooth without troubles.
We would like to notice that an intelligible video
coding for video conferencing can be 100 kbit/sec,
Figure 8: Packet loss of videoconferencing
and FTP over two-hops scatternet
In Figure 8 we show the packet loss rate and the
jitter both during the video conferencing and FTP. The
Gnomemeeting application has a buffer size of 4
seconds to compensate the jitter. With reference of
Figure 6, in the considered scenario the flows cross the
GN/PANU bridge starting from a GN in a piconet and
ending to the PANU in the other piconet (M case) or
starting from a PANU in a piconet and ending to the
GN in the other piconet (S case).
Results show that in case of videoconferencing, the
packet loss is higher than in case of FTP application,
while the jitter size remains quite limited (by the value
of 3 seconds).
As for the perceived quality the audio results
intelligible in all cases. The video presents clear
images in the single link case, while in the two hops
scatternet the quality decreases even if it remains
5.2. J-BlueZ and BASkit performance
A key performance metric for J-BlueZ and BASkit
is the time spent to establish a single link and to create
the whole network. To test J-BlueZ performance we
use 3 nodes and measure the time elapsing from first
module activation to last connection.
We have sequenced node activations with intervals
of 20s, we have chosen 14s between start of
environment inspection and connection trial, 30s
between two automatic inspections. Performance is
strongly biased by these values.
Nodes’ name (A, B and C) reflects the activation
order. Every node executes J-BlueZ, searching for
available devices and then connecting to them. BlueZ
‘scan’ command could return a list with a subset of
available devices. This problem, probably is caused by
inquiry failure [7] and cannot be solved by J-BlueZ.
Therefore J-BlueZ attempts to connect to a reduced
number of nodes compared to available devices. JBlueZ periodically repeats inspection and connection
and eventually connects to newly discovered devices.
At the end we have that all nodes are linked directly to
all the other nodes (fully meshed topology).
Figure 9 presents three examples of formed
networks. As for the time to establish single link
connections we measured: a minimum value of 2s, a
maximum value of 49s and a mean value of 38s.
Measures on the time spent to create network indicate
a minimum value of 78s, a maximum value of 89s and
a mean value of 81s.
J-BlueZ and BASkit aim to different goals to form
network: the former aims at linking a node to every
available device, the latter at connecting a node to the
BASkit has better performance than J-BlueZ since
the topology to be formed is more rational: i.e., BASkit
forms connected network without forming useless
loops as happens in J-BlueZ (right part of Figure 9).
Thanks to the reduced number of links in the best
case, when a single node enters an already formed
network, only one connection is activated. On the other
hand, in the worst case, a large number of isolated
nodes entering concurrently the same area, perform a
reduced number of connection trials since they stop as
soon as the scatternet is connected.
10. Conclusions
This paper presents an experience of design and
performance analysis of an experimental platform
(built up with commercial devices) for Bluetooth adhoc networking. We implemented two suitable tools
(namely J-BlueZ and BASkit) to connect, in a Linux
environment, Bluetooth devices in a scatternet. Some
key applications (data transfer and videoconferencing)
have been tested on the formed network. Scatternet
set-up delay has been also evaluated. Performance
analysis showed the feasibility and simplicity in using
our tools with satisfactory results.
Figure 9: Three J-BlueZ test scenarios
10. References
[1] J. Haartsen, “The Bluetooth Radio System”, IEEE
Personal Communications, Vol. 7, n. 1, pp. 28-36,
February 2000.
[2] P. Johansson, R. Kapoor, M. Gerla, M.
Kazantzidis,“Bluetooth an Enabler of Personal Area
Networking”, IEEE Network, Special Issue on Personal
Area Networks, pp. 28-37, September/October 2001.
[3] T. Salonidis, P. Bhagwat, L. Tassiulas, R. La Maire,
“Distributed topology construction of Bluetooth
personal area networks”, Proc. of the IEEE Infocom
2001, pp. 1577-1586, April 2001.
[4] C. Law, A. Mehta, K-Y Siu, “Performance of a new
Bluetooth scatternet formation protocol”, Proc. of the
Mobihoc 2001.
[5] G. Tan, A. Miu, J. Guttag, H. Balakrishnan, “An
Efficient Scatternet Formation Algorithm for Dynamic
Environments”, in IASTED Communications and
Computer Networks (CCN), Cambridge, November
[6] G. Zaruba, S. Basagni, I. Chlamtac, “BluetreesScatternet formation to enable Bluetooth-based personal
area networks”, Proc. of the ICC 2001, pp. 273-277,
[7] S. Basagni, C. Petrioli, “Multihop Scatternet Formation
for Bluetooth Networks”, Proc. of the VTC 2002, pp.
424-428, May 2002.
[8] F. Cuomo, G. Di Bacco, T. Melodia, “SHAPER: a Self
Healing Algorithm Producing multihop bluetooth
scattERnets”, Proc. of the IEEE Globecom 2003, San
Francisco, pp. 236-240 December 2003.
[9] F. Cuomo, T. Melodia, and Ian F. Akyildiz "Distributed
Self-Healing and Variable Topology Optimization
Algorithms for QoS Provisioning in Scatternets" IEEE
JSAC, Vol. 22, N. 7, pp. 1120-1236 September 2004
[10] Specification of the Bluetooth System v 1.0 B, Volume
1, Core. Bluetooth Special Interest Group, December
[11] Specification of the Bluetooth System v 1.0 B, Volume
2, Profiles, Bluetooth Special Interest Group, December
[12] BlueZ
MANET Experimentations
An Experimental Study of P2P Group-Communication Applications
in Real-World MANETs∗
Franca Delmastro
Andrea Passarella
CNR, IIT Institute
Via G. Moruzzi, 1 – 56124 Pisa, Italy
[email protected]
University of Cambridge, The Computer Laboratory
15 JJ Thomson Avenue – Cambridge CB3 0FD, UK
[email protected]
Group-communication applications are a very promising opportunity for developing valuable MANET-based applications. However, real-world experimental studies are
required to indicate the best solutions to implement them.
We have implemented a real prototype in which alternative networking stacks can be used to support a distributed
Whiteboard application. By means of experimental results
we show that a standard P2P solution based on Pastry and
Scribe is not suitable for MANET environments. We also
show that a cross-layer P2P system optimised for MANETs
(i.e., CrossROAD) is able to overcome many of the problems
experienced with Pastry.
1 Introduction
Even though research on MANETs has been very active in the last decade, real applications addressed to people outside the research community still have to be developed. The typical simulation-based approach for the performance evaluation of MANETs is one of the main reasons of
this. Often, simulation results turn out to be quite unreliable
if compared to real-world measurements [1, 11], and realworld experiments are highly required for MANET applications to become reality, despite their high costs (in terms of
time to set up) and intrinsic limitations (number of nodes).
By leveraging the self-organising nature of MANETs,
group-communication applications can be an outstanding
opportunity from this standpoint. In this paper, we focus on
a significant example of this class of applications, and we
evaluate complete networking solutions that could be used
to develop it. Specifically, we consider the Whiteboard application (WB), which implements a distributed whiteboard
among MANET users. WB allows users to share drawings, messages and other dynamically generated content.
Such a self-organising distributed application can be naturally supported by P2P systems. In our prototype, WB uses
Scribe [4] to share WB content among users via applicationlevel multicast trees. Scribe requires a P2P overlay net∗ This
work was partially funded by the FET-IST Programme of the
European Commission, IST-2001-38113 MOBILE-MAN project.
work based on a DHT. Our prototype includes two alternative P2P solutions, i.e., Pastry [13] and CrossROAD [8].
Both of them provide the same functionalities toward above
layers through the P2P commonAPI [6], but CrossROAD
is explicitly designed for MANET environments. It reduces (with respect to Pastry) the network overhead related
to the overlay management by exploiting cross-layer interactions with a proactive routing protocol. Specifically,
the CrossROAD implementation is compliant to the crosslayer framework described in [7]. Finally, the prototype includes OLSR [12] and AODV [2] at the routing level. Pastry performance is evaluated on top of both routing protocols while CrossROAD is evaluated on top of OLSR, since
CrossROAD is designed to exploit a cross-layer interaction
with a proactive routing protocol.
The main contribution of this paper is evaluating through
real experiments complete networking solutions for developing distributed applications such as WB in real-world
MANETs. We evaluate our prototype at two different levels, i.e., we quantify i) the QoS perceived by WB users, and
ii) the quality of the multicast tree generated by Scribe. First
of all, we show how a proactive routing protocol performs
better than a reactive one with regard to this kind of applications. Then, we highlight that a solution based on Pastry and
Scribe is not suitable for MANET environments. WB users
perceive unacceptable high data loss and delays. Furthermore, both the Pastry overlay network and the Scribe multicast tree get frequently partitioned. This results in some
WB users to be completely isolated from the rest of the network. Finally, we show that some of these problems can
be avoided by using CrossROAD. Specifically, the structure
of the Scribe tree is quite more stable when CrossROAD
is adopted, and partitions problems experienced with Pastry
completely disappear. Thus, CrossROAD turns out to be a
very promising P2P system for MANET environments.
2 WB and its middleware support
The Whiteboard application was originally designed in
[9], and then we adapted it to work on top of Pastry and
CrossROAD. It implements a distributed whiteboard, that
can be used to share dynamically generated content (e.g.,
drawings, messages, . . . ). Each user runs a WB instance on
her mobile device, and selects a topic she wants to associate
to (i.e. “treasure hunting”). Each topic is linked with a canvas on which she can draw strokes or type text. On the same
canvas, the user directly sees strokes and text generated by
others. Being a simple example of group-communication
applications, WB allows us to understand how such applications can be successfully developed in MANETs.
WB needs a subject-based multicast protocol to build
groups (i.e., identify all nodes whose users are interested
into the same topic), and disseminate WB data to the group
members. Specifically, in our testbed, Scribe [4] is used
as the multicast protocol, since it has shown to outperform
other similar solutions [5]. Scribe is designed to work on
top of Pastry, but it can be used with any P2P system providing a commonAPI-compliant overlay network, such as
2.1 Pastry and CrossROAD
Pastry is a P2P system based on a DHT to build a structured overlay network (ring) at the middleware level. A logical identifier (node id) is assigned to each node hashing one
of its physical identifiers (e.g., IP address, hostname). Messages are sent on the overlay by specifying a destination key
k belonging to the logical identifiers’ space. Pastry routes
these messages to the node whose id is numerically closest to k value. To route messages, Pastry nodes maintain
a limited subset of other nodes’ logical ids in their internal
data structures (middleware routing tables). Periodic data
exchange between nodes of the overlay are needed to update the state of the overlay. Finally, in order to initially
join the overlay network, each Pastry node executes a bootstrap procedure, during which it initialises its middleware
routing table by collecting portions of other nodes’ routing
tables. Specifically, each nodes has to connect to an already
existing Pastry node (i.e., it needs to know its IP address) in
order to correctly start the bootstrap procedure.
The bootstrap phase and the periodic data exchange between nodes constitute the main network overhead of Pastry. CrossROAD, that is a Pastry-like P2P system explicitly
designed for MANETs, drastically reduces the Pastry overhead by exploiting cross-layer interactions with a proactive routing protocol. Specifically, CrossROAD defines a
cross-layer Service Discovery protocol in order to broadcast information about upper-layer services (e.g. Scribe)
through the proactive flooding of routing packets, and to
maintain an association between nodes’ IP addresses and
provided services. Hence, each CrossROAD node can autonomously build the overlay, by simply hashing the IP
address of nodes providing the same service. In this way
the overlay network related to a particular service is maintained with almost negligible network overhead in compari-
son with Pastry. Furthermore, CrossROAD i) is completely
self-organising, since it does not require any bootstrap procedure, and ii) correctly manages cases of network partitioning and topology changes with the same delays of the
routing protocols.
2.2 Scribe
Scribe exploits Pastry-like routing to build multicast
groups. From the standpoint of the application running on
Scribe, the group is identified by a topic. Scribe uses the
hash function provided by Pastry (or CrossROAD) to generate the topic id (tid ) in the logical space of node ids. In
order to join the Scribe tree, nodes send a join message on
the overlay with key equal to tid . This message reaches the
next hop (say, N ) towards the destination on the overlay network. The node originating the join message is enrolled
as a child of N . If not already in the tree, N itself joins the
tree by generating a join message anew. Eventually, such
a message reaches the node whose id is the closest one to
tid and is not propagated further. This node is defined as the
root of the Scribe tree.
Application messages are sent on the overlay with key
equal to tid . Hence, they reach the Scribe root, which is
in charge of delivering them over the tree. To this end, it
forwards the messages to its children, which further forward
them to their children, and so on.
Finally, the Scribe maintenance procedure is as follows.
Each parent periodically sends a HeartBeat message to
each child1 . If a child does not receive any message from
the parent for a given time interval (20 s in the default case),
it assumes that the parent has given up, and re-executes the
join procedure. This simple procedure allows node to discover parent failures, and re-join the tree, if the case.
3 Experimental Environment
The experiments reported in this paper are based on a
static MANET. This allows us to highlight limitations that
originate from Pastry and Scribe design, rather than to mobility. Extending the results in the case of mobility is subject
of future work.
The experiment testbed is as depicted in Figure 1. We
set up an indoor MANET consisting of 8 nodes. To have an
homogeneous testbed, all nodes are IBM ThinkPad R50 laptops. We use the built-in Intel PRO-Wireless 2200 802.11
card, with ipw2200 driver (on Linux 2.6 kernel). The data
rate is set to 11 Mbps. In addition the transmission power
of each card has been adjusted to reproduce the topology
shown in the figure and obtain a multi-hop ad hoc network.
During the experiments, nodes marked A through to F participate in the overlay network, and run the WB application (they will be throughout referred to as “WB nodes”).
1 Application-level
messages are used as implicit HeartBeats.
E and bootraps from E. Node A starts 5 seconds after B and
bootraps from B. Finally, node F starts 5 seconds after D and
bootraps from D. After this point in time, the Scribe tree is
created and, finally, WB instances start sending application
messages (herafter, WB messages). This way, the Scribe
tree is built when the overlay network is already stable, and
WB starts sending when the Scribe tree is completely built.
Figure 1. Map of the experiment setup
Nodes marked with “R” are used just as routers. It is worth
pointing out that this setup lies within the “802.11 ad hoc
horizon” envisioned in [11], i.e. 10-20 nodes, and 2-3
hops. Therefore, it is a valid example of possible real-world
In order to have a controllable and reproducible setup, a
human user at a WB node is represented by a software agent
running on the node. During an experiment, each software
agent interleaves active and idle phases. During an active
phase, it draws a burst of strokes on the canvas, which are
sent to all the other WB nodes through Scribe 2 . During an
idle phase, it just receives possible strokes from other WB
nodes. After completing a given number of such cycles (a
cycle is defined as a burst of strokes followed by an idle
time), each agent sends a Close message on the Scribe,
waits for getting Close messages of all the other nodes,
and shuts down. Burst sizes and idle phase lengths are sampled from exponentially distributed random variables. The
average length of idle phases is 10 s, and is fixed through
all the experiments. On the other hand, the average burst
size is defined on a per-experiment basis. As a reference
point, we define a traffic load of 100% as the traffic generated by a user drawing, on average, one stroke per second.
Finally, the number of cycles defining the experiment duration is fixed through all the experiments. Even at the lowest
traffic load taken into consideration, each agent draws – on
average – at least 50 strokes during an experiment. For the
performance figures defined in this paper (see below) this
represents a good trade-off between the experiment duration and the result accuracy.
Some final remarks should be pointed out about the experiment start-up phase. Nodes are synchronised at the beginning of each experiment. Then, in the Pastry case, the
Pastry bootstrap sequence occurs as follows3 : node C starts
first, and generates the ring. Nodes E and D start 5 seconds
after C, and bootstrap from C. Node B starts 5 seconds after
2 Please note that in our experiment each stroke generates a new message to be distributed on the Scribe tree.
3 The same schedule is also used to start CrossROAD, even though a
CrossROAD node does not need to bootstrap from another node.
3.1 Performance Indices
Since Pastry and Scribe have been conceived for fixed
networks, we investigate if they are able to provide an adequate Quality of Service to users in a MANET environment.
To quantify the ”WB user satisfaction” we use two performance indices:
Packet Loss: at each node i, we measure the number of
WB messages received and sent (Ri and Si , respectively) during an experiment; the packet loss experienced by node i is defined as pli = PRiS .
Delay: the time instant when each packet is sent and received is stored at the sending and receiving node, respectively. This way, we are able to evaluate the delay experienced by each node in receiving each packet.
If dij is the delay experienced by node i in receiving
packet j, and Ni the total number of packets received
by i during an experiment, the average
P delay experienced by node i is defined as Di =
Furthermore, we define two more indices, to quantify the
quality of the multicast tree created by Scribe.
Node Stress: for each node, it is defined as the average
number of children of that node. If tij is the time
interval (within an experiment) during which node i
has nj children,
the average node stress of node i is
N S i = Pj
nj tij
Re-subscriptions: for each node, we count the number of
times (during an experiment) this node sends new subscriptions requests, because it can’t communicate with
the previous parent anymore.
4 Performance with Pastry
The results we report in this section are obtained by using Pastry as DHT, and either OLSR or AODV as routing
protocol. Experiments are run by increasing the traffic load
starting from 20% up to 80%.
Before presenting the results in detail, let us define what
herafter will be referred to as “crash of the Scribe Root
Node”. In our configuration Pastry assigns node ids by
hashing the IP address and the port used by Scribe on the
node. Hence, each node always gets the same node id. Furthermore, the topic used by the WB users is always the
Pastry: Packet Loss under Normal Root Behavior
AODV(10%): rings (A),(BCDE),(F)
AODV(20%): rings (A),(BCDEF)
OLSR(20%): rings (ABCDEF)
OLSR(50%): rings (A),(CDEF)
OLSR(80%): rings (BCDE)
Packet Loss (%)
Figure 2. Packet Loss w/o MSRN crash
Pastry: Average Delay under Normal Root Behavior
Avgerage Delay (s)
AODV(10%): rings (A),(BCDE),(F)
AODV(20%): rings (A),(BCDEF)
OLSR(20%): rings (ABCDEF)
OLSR(50%): rings (A),(CDEF)
OLSR(80%): rings (BCDE)
Figure 3. Delay w/o MSRN crash
same. Under the hypothesis that Pastry generates a single
ring encompassing all WB nodes, the Root of the Scribe tree
(i.e., the node whose id is closest to the WB topic id) is the
same through all the experiments, and is node C in Figure 1.
This node will be throughout referred to as the Main Scribe
Root Node (MSRN). Due to the Scribe algorithm, each WB
message to be distributed on the tree is firstly sent to MSRN,
and then forwarded over the tree. Often, this is an excessive load for MSRN, which, after some point in time, becomes unable to deliver all the received messages. Instead,
messages are dropped at the MSRN sending queue. We refer to this event as a crash of MSRN. Of course, since the
application-level traffic is randomly generated, the MSRN
crash is not a deterministic event.
4.1 User Satisfaction
Figures 2 and 3 show the packet loss and the delay indices experienced by the WB nodes considering experiments where the MSRN does not crash. Specifically, we
consider AODV experiments with 10% and 20% traffic
load, and OLSR experiments with 20%, 50% and 80% traffic load, respectively. There is no point in running AODV
experiments with higher traffic load, since performances
with AODV are quite bad, even with such a light traffic
load. In the figure legend we also report the rings that Pastry
builds during the bootstrap phase (please note that, theoretically, just one ring should be built, encompassing all WB
nodes). Finally, an “x” label for a particular node and a particular experiment denotes that for that experiment we are
not able to derive the index related to the node (for example, because some component of the stack crashed during
the experiment).
Figure 2 allows us to highlight an important Pastry weakness. If a WB node is unable to successfully bootstrap, it
starts a new ring, and remains isolated for the rest of the
experiment. In MANET environments, links are typically
unstable, and the event of a WB node failing to contact the
bootstrap node is quite likely. Clearly, once a node is isolated, it is unable to receive (send) WB messages from (to)
other nodes for the rest of the experiment, and this results in
packet losses at all nodes. In the “AODV 10%” experiment,
nodes A and F are isolated, and create their own rings. This
results in packet loss of about 80% at those nodes (i.e., they
just get their own WB messages, which is about one sixth of
the overall WB traffic), and about 33% at nodes B, C, D and
E. Similar remarks apply to the “OLSR 50%” experiment.
It is more interesting to focus on the “AODV 20%” experiment. In this case, node A is isolated, while nodes B, C, D,
E and F belong to the same ring. As before, A’s packet loss
is about 80%. The packet loss at the other nodes due to the
isolation of node A is about 18% (one sixth of the overall
traffic). It is interesting to notice that nodes B and D experience a higher packet loss, meaning that they are unable to
get WB messages generated within the “main” Pastry ring
(i.e., nodes B, C, D, E, F). Finally, in the case “OLSR 20%”,
Pastry is able to correctly generate a single ring, and the
packet loss is quite low. In the case “OLSR 80%” nodes A
and F crash. However, the packet loss experienced by the
other nodes is negligible.
Similar observations can be drawn by focusing on the
delay index (Figure 3). First of all, it should be pointed out
that the delay related to nodes that are the sole member of
their own ring (e.g., node A in the “AODV 10%” case) is obviously negligible. Even though – in general – the delay in
this set of experiments is low, it can be noted that better performances are achieved by using OLSR instead of AODV.
Finally, it should be noted that MSRN (node C) always experiences a lower delay with respect to the other nodes in
the same ring.
Figures 4 and 5 show the packet loss and the delay indices in cases of MSRN crash. The packet loss experienced
by nodes in the same ring becomes higher than in cases
where MSRN does not crash. In the first three experiments,
node A isolation causes a packet loss of about 18% on the
other nodes. Hence, the remaining 60% packet loss is ascribed to the MSRN crash. Quite surprisingly, OLSR with
80% traffic load shows better performance than OLSR with
50% traffic load. It is also interesting to note that the packet
loss at MSRN is always lower than at other nodes in the
same ring. This highlights that MSRN is able to get, but un-
Pastry: Packet Loss under Root Crash
Pastry: Node Stress under Normal Root Behavior
AODV(10%): rings (A),(BCDEF)
AODV(20%): rings (A),(BCDEF)
OLSR(50%): rings (A),(CDEF)
OLSR(80%): rings (ABCDEF)
Avgerage Node Stress
Packet Loss (%)
AODV(10%): rings (A),(BCDE),(F)
AODV(20%): rings (A),(BCDEF)
OLSR(20%): rings (ABCDEF)
OLSR(50%): rings (A),(CDEF)
OLSR(80%): rings (BCDE)
Figure 6. Node stress w/o MSRN crash
Pastry: Node Stress under Root Crash
Pastry: Average Delay under Root Crash
Figure 4. Packet Loss w/ MSRN crash
AODV(10%): rings (A),(BCDEF)
AODV(20%): rings (A),(BCDEF)
OLSR(50%): rings (A),(CDEF)
OLSR(80%): rings (ABCDEF)
AODV(10%): rings (A),(BCDEF)
AODV(20%): rings (A),(BCDEF)
OLSR(50%): rings (A),(CDEF)
OLSR(80%): rings (ABCDEF)
Avgerage Node Stress
Avgerage Delay (s)
Figure 5. Delay w/ MSRN crash
able to deliver over the Scribe tree WB messages generated
by other nodes. Similar observations can be drawn by looking at Figure 5, as well. The delay experienced by nodes B,
D, E and F can be as high as a few minutes, either by using
AODV or OLSR. Finally, the delay experienced by MSRN
is very low in comparison to the delay experienced by the
other nodes.
To summarise, the above analysis allows us to draw the
following observations. The Pastry bootstrap algorithm is
too weak to work well in MANETs, and produces unrecoverable partitions of the overlay network. This behavior is generally exacerbated by AODV (in comparison to
OLSR). Furthermore, MSRN is clearly a bottleneck for
Scribe. MSRN may be unable to deliver WB messages
also with moderate traffic loads, resulting in extremely high
packet loss and delay. Moreover, the performance of the
system in terms of packet loss and delay is unpredictable.
With the same protocols and traffic load (e.g., OLSR and
50% traffic load), MSRN may crash or may not, resulting
in completely different performance figures. In cases where
MSRN crashes, packet loss and delay are clearly too high
for WB to be actually used by real users. However, even
when MSRN does not crash, the high probability of WB
users to be isolated from the overlay network makes Pastrybased solutions too unreliable. These results suggest that
Pastry and Scribe need to be highly improved to actually
support group communication applications such as WB in
Figure 7. Node stress w/ MSRN crash
MANET environments.
4.2 Multicast Tree Quality
In this section we analyse the node stress and resubscription indices, with respect to the same experiments
used in the previous section.
Figures 6 and 7 plot the average node stress with and
without MSRN crashes, respectively. In both cases, the
node stress is significantly higher at MSRN than at any
other node. This means that the Scribe tree is a one-level
tree, and MSRN is the parent of all the other nodes. This
behavior is expected, and can be explained by recalling
the way Scribe works. In our moderate-scale MANET, all
nodes are in the Pastry routing table of each other. Hence,
Scribe join messages reach MSRN as the first hop, and
MSRN becomes the parent of all other nodes (in the same
ring). Together with the way application-level messages are
delivered, this phenomenon explains why MSRN is a bottleneck, since it has to send a distinct message to each child
when delivering WB messages over the tree. This is a major
limitation of the Scribe algorithm, and optimisations of the
P2P system are clearly not sufficient to cope with it.
In Figures 6 and 7 we have added “R” labels to indicate
nodes that occur to become Scribe Root during the corresponding experiment. When MSRN does not crash (Figure 6) other nodes become Scribe root only as a side effect of a failed Pastry bootstrap. On an isolated WB node,
Pastry: Re-subscriptions under Normal Root Behavior
Number of Re-subscriptions
AODV(10%): rings (A),(BCDE),(F)
AODV(20%): rings (A),(BCDEF)
OLSR(20%): rings (ABCDEF)
OLSR(50%): rings (A),(CDEF)
OLSR(80%): rings (BCDE)
Figure 8. Re-subscriptions w/o MSRN crash
Pastry: Re-subscriptions under Root Crash
Number of Re-subscriptions
AODV(10%): rings (A),(BCDEF)
AODV(20%): rings (A),(BCDEF)
OLSR(50%): rings (A),(CDEF)
OLSR(80%): rings (ABCDEF)
rect view of the network at the Pastry level, originated from
Figures 8 and 9 show the re-subscription index for the
same set of experiments. Figure 8 shows that, when MSRN
does not crash, the Scribe tree is quite stable. Most of the
re-subscriptions occur at node F, which is the “less connected” node in the network (see Figure 1). In these experiments, the performance in the AODV cases is worse than in
OLSR cases. Furthermore, upon MSRN crashes (Figure 9),
the number of re-subscriptions increases drastically, even
in case of “well-connected nodes” (i.e., node B, D and E).
MSRN crashes make other nodes unable to get messages
from their parent (i.e., MSRN itself), increasing the number
of re-subscriptions. It is interesting to point out that this is
a typical positive-feedback control loop: the more MSRN
is congested, the more re-subscriptions are sent, the more
congestion is generated.
To summarise, the multicast tree generated by Scribe on
top of Pastry is quite unstable, especially in cases of MSRN
crashes. The tree may get partitioned in disjoint sub-trees,
and many re-subscriptions are generated by nodes. Furthermore, Scribe is not able to generate a well-balanced multicast tree, since MSRN is the parent of all other nodes. Directions to optimise Scribe are discussed in Section 6.
Figure 9. Re-subscriptions w/ MSRN crash
Scribe builds a tree which consists only of the node itself,
that is thus the root. However, Scribe partitions may also
occur due to congestion at the Pastry level in cases where
MSRN crashes. By looking at Figure 7, it can be noticed
that nodes other than MSRN may become root also if they
belonged (after the Pastry bootstrap phase) to the same overlay network of MSRN. This phenomenon occurs, for example, at node A in the OLSR 80% case, and at node B and
F (whenever they become root). It should be noted that a
node with id n1 (other than MSRN) becomes root when i)
it looses its previous parent, and ii) the Pastry routing table
does not contain another node id n2 such that n2 is closer
to the WB topic id than n1 . Figure 7 shows that the congestion at the Pastry level is so high that the Pastry routing
table of some nodes becomes incomplete (i.e., MSRN disappears from other nodes’ routing table). Thus, the Scribe
tree gets partitioned in several isolated sub-trees. Clearly,
this contributes to the high packet loss measured in these
experiments. Another effect of Pastry congestion during
MSRN crashes is a possible reshaping of the Scribe tree.
Figure 7 shows that the average Node Stress of E is close
to 1 in the “AODV 20%” and “OLSR 80%” cases. This
means that MSRN disappears from the Pastry routing table
of some node, which – instead of becoming a new root –
finds node E to be the closest one to the WB topic id. This
phenomenon could be considered a benefit, since it reduces
the MSRN node stress. However, it derives from an incor-
5 Improvements with CrossROAD
In this section we show that using a P2P system optimised for MANETs is highly beneficial to the stability of
the Scribe tree. In this set of experiments, we use CrossROAD instead of Pastry, and set the traffic load to 20%,
50% and 100%, respectively. We concentrate on the performance figures related to the quality of the multicast tree,
i.e., the average node stress (Figure 10) and the number of
re-subscriptions (Figure 11). A complete evaluation of the
User Satisfaction parameters, as well as further optimisations of the Scribe algorithm, are subjects of future work.
The first main improvement achieved by using CrossROAD is that neither the overlay network nor the Scribe
tree get partitioned. CrossROAD is able to build a single
overlay network in all the experiments. Furthermore, even
at very high traffic loads (e.g., 100%), MSRN is the only
root of the Scribe tree. Therefore, CrossROAD is able to
overcome all the partition problems experienced when Pastry is used.
Figure 10 clearly shows that the node stress still remains
quite unbalanced among the nodes. MSRN is typically the
parent of all other nodes, and this contributes to make it a
bottleneck of the system, as highlighted above. This behavior is expected, since it derives from the Scribe algorithm,
and cannot be modified by changing P2P system.
Finally, Figure 11 shows that the Scribe tree is more stable (i.e., requires less re-subscriptions) using CrossROAD
instead of Pastry. To be fair, we have to compare Fig-
CrossRoad: Node Stress for increasing loads
20%: ring (ABCDEF)
50%: ring (ABCDEF)
100%: ring (ABCDEF)
Average Node Stress
Figure 10. Node Stress with CrossROAD
CrossRoad: Re-subscriptions for increasing loads
Number of Re-subscriptions
20%: ring (ABCDEF)
50%: ring (ABCDEF)
100%: ring (ABCDEF)
Figure 11. Re-subscriptions with CrossROAD
ure 11 with both Figures 8 and 9. It is clear that CrossROAD outperforms Pastry when used on top of AODV. The
“20%” case of CrossROAD should be compared with the
“OLSR 20%” case of Figure 8, since in both experiments
the overlay network is made up of all nodes. The number of re-subscriptions measured at node F is the same in
both cases, while it is higher at node E when Pastry is
used. The CrossROAD “50%” case shows a higher number of re-subscriptions with respect to the “OLSR 50%”
case in Figure 8. However, it should be noted that in the
latter case the overlay network encompasses less nodes,
and hence the congestion is lower. It should also be noted
that, with the same nodes in the overlay network, with the
same protocol stack and traffic load, Pastry experiments
may suffer MSRN crashes (Figure 9). In this case, the number of re-subscriptions is much higher than in the CrossROAD case. Finally, results in the CrossROAD “100%”
case should be compared with the “OLSR 80%” case of
Figure 9, since the overlay network is the same in both experiments. CrossROAD achieves comparable performance,
and at some nodes it outperforms Pastry, even if the application traffic is significantly higher.
5.1 Overlay management overhead
In the previous section we have shown that adopting CrossROAD significantly improves the performance of
Scribe. In this section we highlight that one of the main
reasons for this improvement is the big reduction of the network overhead. This is a key advantage in MANET environments.
Figure 12 shows the network load experienced by nodes
A, C and by the two nodes which just act as routers, during the Pastry “OLSR 80%” experiment in which MSRN
crashes4 . Each point in the plot is computed as the aggregate throughput (in the sending and receiving directions)
over the previous 5-seconds time frame. We take into consideration the traffic related to the whole network stack,
from the routing up to the application layer. Specifically,
nodes A and C are representative for WB nodes, pointing
out the difference with nodes that just work as routers. The
discrepancy between the curves related to node A and C
confirms that the MSRN node has to handle a far greater
amount of traffic with respect to the other WB nodes, due
to the Scribe mechanisms. Furthermore, it should be noted
that the curves related to the two routers can hardly been
distinguished in Figure 12, since they are about 400Bps.
This means that the lion’s share of the load on WB nodes is
related to Pastry, Scribe and the WB application.
Figure 13 plots the same curves, but related to the
“100%” CrossROAD experiment. Also in this case, MSRN
(node C) is more loaded than the other WB nodes. However,
by comparing Figures 13 and 12 we can highlight that the
Pastry network load is far higher than the CrossROAD network load. By considering the average value over all nodes
in the MANET, the Pastry load is about 3 times greater than
the CrossROAD load. More specifically, the average load
of C and A is 48.5 KB/s and 16.5 KB/s in the Pastry case,
while drops to 21.1 KB/s and 2.96 KB/s in the CrossROAD
case. The reduction of the network load achieved by CrossROAD is thus 56% at node C and 82% at node A. Since the
other stack components are exactly the same, CrossROAD
is responsible for this reduction 5. Furthermore, it should be
noted that, during several time intervals, the load of node
A is just slightly higher than that of “routing” nodes. This
suggests that the additional load of CrossROAD management with respect to the routing protocol is very limited.
6 Conclusions and Future Works
Results presented in this paper allows us to draw the
following conclusions. Pastry and Scribe seem not to be
good candidates to support group communication applications in MANET environments. Pastry is particularly weak
during the bootstrap phase, causing the overlay network to
be partitioned into several subnetworks, and some nodes to
be unable to join application services. Further partitions
4 We do not take into account AODV experiments, since OLSR has
clearly shown to outperform AODV.
5 The actual reduction is even higher, since the application-level traffic
is 100% in the CrossROAD case.
MSRN, messages can be duplicated and delivered at each
branching point in the tree. These policies are expected to
drastically mitigate the bottleneck problems experienced by
Scribe. Furthermore, they raise very interesting arguments
about causal ordering of messages, that we are planning to
address, as well.
WB on Pastry
Network Load (B/s)
Figure 12. Network Load with Pastry
WB on CrossRoad
Network Load (B/s)
Figure 13. Network Load with CrossROAD
may occur in the Scribe tree due to congestion at the Pastry level. Finally, the delivery algorithm implemented by
Scribe generates a severe bottleneck in the tree, which is
highly prone to get overladed. All these limitations result
in unacceptable levels of packet loss and delay for applications. Many of these problems can be avoided by adopting
a cross-layer optimised P2P system such as CrossROAD.
Thanks to the interactions with a proactive routing protocol
CrossROAD is able to avoid all the partition problems experienced with Pastry, and to drastically reduce the network
overhead. Clearly, CrossROAD cannot solve the problem
of bottlenecks in the Scribe trees. Therefore, optimised versions of Scribe are required for group communication applications such as WB to be really developed in MANETs.
The direction we are exploring is building a single distribution tree, optimised through cross-layer interactions with a
proactive routing protocol. Building a single-tree, instead
of a new tree for each source node, allows for a more scalable solution. In addition, cross-layering allow us to retain
the subject-based features of Scribe (e.g., locating a tree by
means of its topic), while exploiting also topological information to build the tree. Furthermore, the tree can be built
in a completely distributed way, by exploiting greedy policies such as those used in YAM [3] and ALMA [10]. Finally, the data-distribution phase can be optimised so as to
avoid each message to be sent to the MSRN and then delivered to other nodes. For example, while travelling towards
[1] G. Anastasi, E. Borgia, M. Conti, E. Gregori and A.
Passarella, “Understanding the Real Behavior of Mote and
802.11 Ad hoc Networks: an Experimental Approach”,
Pervasive and Mobile Computing, in press.
[2] AODV, Dept. of Information technology at Uppsala University (Sweden), henrikl/aodv/.
[3] K. Carlber and J. Crowcroft, “Building Shared Trees Using
a One-to-Many Joining Mechanism”, ACM Computer
Communication Review, pp. 5-11, Jan. 1997.
[4] M. Castro, P. Druschel, A-M. Kermarrec and A. Rowstron,
“SCRIBE: A large-scale and decentralised application-level
multicast infrastructure”, IEEE Journal on Selected Areas
in Communication (JSAC), Vol. 20, No, 8, October 2002.
[5] M. Castro, M. B. Jones, A-M. Kermarrec, A. Rowstron, M.
Theimer, H. Wang and A. Wolman, “An Evaluation of Scalable Application-level Multicast Built Using Peer-to-peer
overlays”, Infocom 2003, San Francisco, CA, April, 2003.
[6] F. Dabek and B. Zhao and P. Druschel and J. Kubiatowicz
and I. Stoica, “Towards a common API for Structured
Peer-to-Peer Overlays”, Proc. of the the 2nd International
Workshop on Peer-to-peer Systems (IPTPS’03), Berkeley,
CA, Feb. 2003.
[7] M. Conti, G. Maselli, G. Turi, and S. Giordano, “Cross
layering in mobile ad hoc network design”, IEEE Computer,
Feb. 2004.
[8] F. Delmastro, “From Pastry to CrossROAD: Cross-layer
Ring Overlay for Ad hoc networks”, in Proc. of Workshop
of Mobile Peer-to-Peer 2005, in conjuction with the PerCom
2005 conference, Kauai Island, Hawaii, Mar. 2005.
[9] M. Dischinger, “A flexible and scalable peer-to-peer multicast application using Bamboo”, Report of the University
of Cambridge Computer Laboratory, 2004, available at
[10] M. Ge, S.V. Krishnamurthy, and M. Faloutsos, “Overlay
Multicasting for Ad Hoc Networks”, Proc. of the Third
Annual Mediterranean Ad Hoc Networking Workshop
(MedHocNet 2004), June 2004.
[11] P. Gunningberg and H. Lundgren and E. Nordström and
C. Tschudin, “Lessons from Experimental MANET Research”, Ad Hoc Networks Journal, (Special Issue on “Ad
Hoc Networking for Pervasive Systems”), Vol. 3, Number
2, March 2005.
[12] OLSR, Andreas Tonnesen, Institute for informatics at the
University of Oslo (Norway),
[13] A. Rowstron and P. Druschel, “Pastry: Scalable, distributed
object location and routing for large-scale peer-to-peer
systems”, Middleware 2001, Germany, November 2001.
A Comparative Study of Cooperative Algorithms for Wireless Ad Hoc Networks
Alan Lim, Vikram Srinivasan, Chen-Khong Tham
Department of Electrical and Computer Engineering
National University of Singapore
Engineering Drive 3 Singapore 117576
{eng10346, elevs, eletck}
In Ad-Hoc Networks (AHN), nodes can take on the role
of a source, destination or relay. Relays are needed when
source nodes wish to communicate with destination nodes
that are far away. However, in acting as relays, nodes
expend a significant amount of energy transmitting data
across the network. Energy is a precious commodity for
an AHN and the aim of nodes is to maximize their throughput and survive for the longest period possible. Thus, nodes
would not be willing to act as relays all the time. As a result, efficient cooperative algorithms are needed in AHNs
to determine when a node should act as a relay, thus ensuring the survivability and performance of nodes. However,
there are issues that need to be considered before existing
cooperative algorithms may be implemented in an actual
AHN. This paper focuses on the research, modification, implementation, verification and analysis of the Generous Tit
for Tat (GTFT) and Nuglet cooperative routing algorithms
on an actual AHN.
1. Introduction
Ad-hoc Networks (AHNs) have vast potential applications in the modern world. These include situations that
demand the rapid deployment of nodes. AHNs can also
extend the coverage of a network significantly without incurring significant financial costs. A node in an AHN can
typically take the form of a PDA or a laptop. AHNs usually have an arbitrary and random configuration. In fact, the
number of nodes that come together to form a network at
any instance is non-deterministic. Furthermore, we assume
that nodes have an equal chance at any one time of being a
source, destination or relay.
A node in an AHN is faced with two primary constraints.
Firstly, in the transmission of data packets, energy (in terms
of the battery levels of the nodes) is consumed. Thus,
since PDAs or laptops are portable systems allowing users
to process information on the go, they are heavily dependant on the limited battery power that they carry. Consequently, nodes would want to conserve as much energy as
possible. Secondly, nodes would also like to have the maximum goodput (number of packets accepted by the relays
over the number of packets sent out as a source) possible.
However this would require relay nodes to cooperate all the
time. While it may be intuitive that relay nodes should help
to forward packets for other nodes all the time, it is not in
their interest to do so. If a relay node were to transmit data
continuously for other nodes, there may be little or no energy left for its own use.
References such as [1], [2], [3], [4] and [5] propose the
use of cooperative algorithms to determine if a node should
forward a data packet for another node. The primary aims
of these algorithms are to maximize the throughput and lifetime of a node. Other approaches like [6], [7] and [8] propose mechanisms to be employed in the event of misbehaving nodes. Meanwhile, references [9] and [10] have studied
the energy consumption of wireless transmission.
This paper uses some of these ideas with slight modifications to the theoretical framework where necessary. While
previous works have been supported purely by simulation
results, this is, to the best of the authors’ knowledge, the first
paper that looks at the implementation issues of cooperative
algorithms on an actual AHN. Specifically, the GTFT [1]
and Nuglet [2] cooperative algorithms were implemented.
In this paper we study the results obtained and verify them
against existing simulation results. We also compare the
algorithms against each other and with the case when no
cooperative algorithm was used.
In section 2, the algorithms employed in [1] and [2] are
briefly discussed. Next, in section 3, the setup of the experiments is given. In section 4, the necessary modifications
to the existing Nuglet and GTFT algorithms are discussed
in detail and their implementation is explained. After that,
in section 5 the experimental settings are given, while the
experiment results are shown and analyzed in section 6. Finally, in section 7, the paper is concluded.
2. Related Work
2.1. Nuglet Algorithm
The Nuglet algorithm, described in [2] seek to stimulate packet forwarding by rewarding nodes who participate
in packet forwarding. It does this through the use of a
nuglet counter, which is incremented whenever nodes forward packets for others. Conversely, the counter is decremented by an integer constant N whenever a node acting
as a source, sends out a packet. Nodes are not allowed
to send out packets if they do not have sufficient counters.
Hence, nodes are encouraged to forward packets in order to
increase their counter values, thereby enabling them to send
out data packets subsequently. Besides the nuglet counter,
the algorithm also includes a battery counter. The battery
counter is decremented whenever a node sends or receives
a data packet. It represents the number of packets that a
node can send out before its battery runs out. In [2], the
routing decisions of a node is found to be constrained by
the following equations.
W here :
Outr , Outs ≥ 0
N ∗ Outs − Outr ≤ C
Outs + Outr = B
: N umber of packets sent out by the node
: N umber of packets relayed by the node
: Integer constant
: N uglet counter
: Battery counter
From equations (1) - (3), Fig 1. is obtained. The optimum operating point for the nodes corresponds to the intersection between the two lines in Fig 1. The maximum number of output and forwarded packets in equations (4) and (5)
are then obtained from the respective Outs and Outr values at the intersection point. When a relay has forwarded as
many packets as the maximum number of output packets,
further relay requests will not be accepted. This is because
its not in it’s interest to forward any more packets, as it has
acquired enough nuglets to send out as many packets as a
source to drain out its battery counter.
Maximum no. of output packets
= B+C
N +1
Maximum no. of forwarded packets
While the Nuglet algorithm performs packet based processing like other micro payment schemes. It does not require
any additional infrastructure (such as the accounting center
or base stations described in [4]) that makes the realisation
Figure 1. Maximisation of Source Output under Nuglet Algorithm from [2]
of an AHN tricky. Furthermore, unlike other similar micro
payment schemes such as [4], it does not involve additional
information such as the ticket data in every packet. Thus
with a smaller packet size, fewer network congestions are
expected. Bearing these in mind, it has the potential to serve
as a simple and efficient solution.
The work on the Nuglet algorithm in [2] focuses primarily on how nodes can be stimulated to forward packets. It
assumes that without cooperative algorithms in place, none
of the nodes will be inclined to forward packets. It shows
how the throughput of the nodes can be improved, when
nodes are stimulated to forward packets with the expectation that other nodes will return the favor. The paper however does not address the effects of the algorithm on the
overall performance of the network in terms of energy consumption and potential energy savings. There is no comparison made between the network before and after the employment of the Nuglet algorithm.
2.2. GTFT Algorithm
The GTFT algorithm is based on game theory as described in economics. Unlike other routing algorithms including the Nuglet algorithm, nodes under GTFT are not
rewarded with virtual currency whenever they forward a
data packet. Instead, they are encouraged to forward a data
packet because they expect other nodes to react negatively
and drop their own requests in the future if they do not cooperate by forwarding packets as often as they should.
Another difference is that GTFT is a session based routing algorithm. Before a source node sends out any data
packets, it will first send out a session request. Relays then
decide whether to accept all relay requests for the entire session or none at all. Nodes refuse session requests either if it
has forwarded more sessions than the Pareto optimal value
or if the rate at which it is accepting sessions (less the generosity value, ε) is greater than the rate at which other nodes
are accepting its session requests. In other words, if the generosity value is increased, nodes are more willing to accept
session requests even if they have not received an equivalent
amount of help.
Next, nodes are classified according to their power constraints. Power constraint is defined to be the ratio between
the initial energy allocated to the node and its expected lifetime. The corresponding power constraint associated with
the current session can then be derived from the node with
the lowest power constraint. This is because there is no
incentive for a node with a higher power constraint to behave more liberally in forwarding packets since it knows
that other nodes with lower power constraints are not able
to reciprocate in kind.
For a particular session, a node can be a source, relay or
destination. Energy is spent by a node both as a source as
well as a relay for others. In fact, the total amount of energy
that is spent as a source as well as a relay will be constrained
by its power constraint. This is illustrated in (6).
espj + erpj ≤ pclass(p)
W here :
esj :
erj :
pclass(p) :
Average energy spent per slot as a source
Average energy spent per slot as a relay
T otal number of sessions
Current session
P ower constraint of particular class p
GTFT also has the feature of a dynamic algorithm. The
forwarding rate is not only dependent on the initial power
settings but on the prevailing network condition as well. If
the current rates at which other nodes are accepting relay
requests are less than what the node itself is accepting for
instance, the node will adapt accordingly and forward less
packets for others in the future.
In light of the features described, GTFT has the potential
to serve as an efficient solution just like the Nuglet algorithm. However, while the Nuglet algorithm adopts a packet
based approach, GTFT performs its routing decision at a
session level. The dynamic nature of GTFT also contrasts
with the relatively static Nuglet algorithm. The two algorithms are thus chosen for study because they represent two
plausible yet different solutions.
3. Experimental Setup
This section details the setup of the nodes used in the
experiments. Details of the physical setup is given in Table
1. A single relay setup proved to be adequate in capturing
Table 1. System Specifications
Hardware 1 Toshiba Tecra 9100 laptop
1 Compaq iPAQ 3630
1 HP iPAQ 5500 Pocket PC
802.11b Cisco Aironet 350 wireless card
Operating iPAQs:
Familiar project Linux v0.7.2 distribution
with a 2.4.29-rmk6-pxa1-hh30 kernel
Redhat 9.0 with a 2.4.20-8 kernel
Topology 3 node string topology
the salient points of our research without involving a large
number of operators and nodes. The iPAQ 5500 came with
its own wireless LAN support while the Cisco Aironet 350
wireless card were used by the laptop and the iPAQ 3630.
3.1. Packet Forwarding
The physical implementation of an actual AHN can
prove to be fairly difficult. Firstly, nodes have to be placed
a significant distance apart in order for a hop to occur. Secondly, the high mobility of nodes in an AHN means that
nodes have to constantly change their positions with reference to each other. These features means that a large physical space is needed to accommodate all the nodes. Furthermore, operators might be needed at each node to change the
geographical positions of the nodes during the experiment.
To overcome the first constraint, we modified the way
routing tables were established. Specifically, the changes
were made such that end nodes can only accept route replies
from relay nodes and not from each other during route discovery. Unlike end nodes, the routing table of the relay
node remains unchanged, and it was able to receive route
replies from all it’s neighbors during route discovery. These
changes allowed all the nodes to be placed next to each other
during the experiment and ensured that all packet transfers
went through an intermediate node.
To overcome the second constraint, we had a command
center that randomly selected nodes at the start of each session to take on either the roles of a source, relay or destination. For a 3 node topology, there were 6 possible configuration that resulted from this selection. After the roles
were defined, the information was broadcasted to the nodes.
Nodes then configured their routing tables based on this information. After the nodes have been configured, the source
node proceeded to send out the data packets to the destination node. At the end of the session, the command center
was notified so that initialization and configuration was carried out for the next test case.
3.2. Measuring Energy Consumption
In [9] and [10], energy consumption was measured directly from the voltage and current drawn by the network interface card (NIC). This required an external piece of hardware such as the Sycard CardBus extender that allowed the
NIC to be inserted into it. The energy measurements were
then taken via a digital oscilloscope and multimeter to the
terminals of the extender.
It is not in line with the objectives of this paper to model
the energy consumption of data packets. Instead, we applied the average current and voltage measurement in [10],
to obtain the average energy consumed in sending, relaying
and receiving packets. The power constraint and initial battery counter values in our experiments were obtained this
way as well.
This approach has several advantages. Firstly, nodes
could be plugged into a power supply permanently during
the course of the experiment instead of running on batteries.
Without recharging the batteries after every experiment, a
considerable amount of time is saved. Next, a fairer measurement is achieved as the amount of energy consumed
in transmitting data packets is constant in all experiments.
Thus, the difference in results obtained in the experiment is
purely due to the routing algorithms and not the result of any
inconsistency that may arise during the course of measuring
the energy consumption physically. Finally, the results obtained can be easily applied to other nodes with different
power specifications by substituting their individual specifications into the equations.
4. Modifications made to existing algorithms
4.1. Nuglet Algorithm
In [2], the energy consumed in sending out packets as a
source and as a relay is assumed to be constant and equal to
1. Thus, the battery counter is chosen to be equal to the sum
of the number of forwarded and generated packets.
4.1.1 Inclusion of packet received term
However, papers such as [9] and [10] indicates that the
amount of energy consumed in receiving packets is significant and cannot be ignored. Thus, the term Outd should be
added to the equations. Outd here refers to the number of
packets that a node would like to receive as a destination.
Outd , Outs , Outr ≥ 0
N ∗ Outs − Outr ≤ C
Equation (7) involves an additional Outd term while
Equation (8) is the same as the original equation in [2]. The
term Outd has no effect on the nuglet counter here as the
cost of transmitting the packet had already been paid by the
source node. Thus, the only effect of receiving a data packet
is to decrement the battery counter. Equation (9) shows the
maximum possible energy consumption of the node from
the individual roles given the limited energy available.
Maximum possible power consumption
= Erecv (Outd ) + Esource (Outs ) + Erelay (Outr )
W here :
: Energy consumed in receiving a data packet
: Energy consumed in relaying a data packet
: Energy consumed in sending a data packet
4.1.2 Difference in energy consumption
Although the energy consumption of nodes varies according to roles in the real system, the difference between the
energy consumed as a source and relay is minimal. They
are therefore treated equally here. The energy consumed as
a destination and the energy consumed as a source or relay can be expressed as a ratio. This simplifies calculations
later in the paper.
Energy consumed as a destination
Energy consumed as a source/relay
In [2], the value of the battery counter B, is equal to the
number of packets that a node can send out as a source.
Similarly here, the value of B can be obtained by dividing the total energy available by Esource . This gives the
maximum number of packets that a node can send out as a
source. This is shown in Equation (10).
Maximum possible power consumption
= Erecv (Outd ) + Esource (Outs ) + Erelay (Outr )
= Erecv (Outd ) + Esource (Outs + Outr )
= Esource (r∗ Outd + Outs + Outr )
Total no. of packets it is able to send out as a source
= B = (r∗ Outd + Outs + Outr )
Using the equations (7), (8) and (11), we plot Fig. 2.
The optimum operating point is again given to be the intersection between the two lines. Based on this graph, we obtain the maximum number of output packets and forwarded
packets as follows:
Maximum no. of output packets
d +C
= B−r NOut
Maximum no. of forwarded packets
Outd )−C
= N (B−rN +1
Table 2. Experimental Settings
Number of nodes
Number of test cases
Size of payload (bytes)
Number of packets per test case 1,000
Number of sessions per test case 1
Configuration time interval (s)
Routing algorithm
Packet generation rate
Figure 2. Modified maximization of output
packets graph under Nuglet algorithm
instead of looking at the power constraint, we associate the
energy class with τ , the probability of a node accepting a
relay request. A higher probability of accepting a relay request means a higher energy class.
4.2. GTFT Algorithm
In [1], the energy consumed as a destination is also assumed to be negligible and the energy required in relaying
a session is taken to be constant and equal to 1 for all nodes.
However, as in the Nuglet scheme, modifications has to be
made to the existing equations.
4.2.1 Additional Destination Term
The energy consumed in receiving packets needs to be considered and the modified power constraint equation with the
additional receiving energy term should then be
espj + edpj + erpj ≤ pclass(p)
W here :
esj :
erj :
edj :
pclass(p) :
Average energy spent per slot as a source
Average energy spent per slot as a relay
Average energy spent per slot as a destination
T otal number of sessions
Current session
P ower constraint of particular class p
4.2.2 Factoring the Average Energy Consumed
Next, the energy consumed in data transmission depends on
the hardware architecture of the nodes and is not equal to 1.
Factoring in the difference in energy consumed per packet
for different roles affects the way energy classes are determined. Unlike in [1] where energy classes are determined
simply by looking at the power constraint of the nodes, there
is a need to consider the average energy consumed by the
node in each role as well. iPAQs, for example, are known to
be more energy efficient. For each packet transmission, less
energy is consumed compared to that for a laptop. Thus,
5. Experimental Settings
In Table 2, the experiment settings are outlined. A configuration time interval was introduced between sessions
because nodes need time after each session to prepare their
routing table for the next test case. The AODV routing algorithm is chosen for its ability to perform well under situations which demand a higher level of mobility [11].
The cooperative algorithms tested here were deployed
with slight modifications to the kernel module developed by
NIST [12]. Besides the cooperative algorithms mentioned
earlier, experiments were also conducted using the original
kernel module without modifications. We will refer to this
in our discussion as the Original AODV scheme.
In [2], the nodes are assumed to be generating packets at
a fixed rate. We have however, taken a different approach
in our experiments. In our experiments, we assumed that
a single session was carried out for each test case. Before
the start of each test case, the nodes were randomly chosen
to assume the role of a source, relay or destination. The
rates at which the individual nodes were generating packets
was thus random and dependent on the frequency it was
chosen to be a source. This approach adds realism to the
experiments, as nodes in an actual network do not have a
fixed role, and their roles are expected to change over time.
The measurement of energy consumption was restricted
to the duration of the session as the energy consumed during
the configuration time interval is inconsequential. In reality,
the roles that nodes plays in a particular session are defined
by the nodes themselves instead of an external command
center. The concept of a configuration time interval then,
would be irrelevant.
Finally, we assumed that nodes in the AHN are not malicious and do not misbehave. Consequently, we do not consider measures such as the watchdog and pathrater in [6],
CONFIDANT in [7] and CORE in [8].
6. Experiment Results
6.1. Comparison between Routing Algorithms
In an AHN, the performance of a node is affected by
other nodes. Hence, improving the performance of one
node could be at the expense of others. However, as long as
the overall performance of the network has improved, the
optimisation objectives have been met. Here, the respective
algorithms are compared against several performance indicators.
Figure 3. Portion of energy spent relaying under the different schemes for the individual
nodes and the overall AHN.
6.1.1 Relay Ratio
Ideally, a node would want as low a relay ratio (proportion
of packets that are relayed for others as compared to the total number of incoming and outgoing packets) as possible.
In the experiment, we observe that while the relay ratio has
improved for some of the nodes under the GTFT scheme,
this ratio has deteriorated for the other nodes. However,
when the Nuglet scheme was adopted, a relay ratio of 33%
was achieved by all nodes. This suggests that the Nuglet
algorithm would lead to a more equitable forwarding burden between nodes. The unequal distribution under Original AODV and GTFT is attributed to the different packet
generation rate among nodes. In our experiment, the rate
at which individual nodes were chosen to be a source node
was random and thus the number of relay requests varied
between nodes. This is a realistic view of how the network
would function in the short run. For a particular time interval, it is reasonable to expect that some nodes will have
more requests than others. However in the long run, we can
expect all nodes to have the same number of relay requests
and the disparity described earlier would no longer by relevant. Thus, the advantage that the Nuglet algorithm has
in ensuring that the forwarding burden is evenly distributed
might not be as evident in the long run.
6.1.2 Portion of Energy Spent As a Relay for Others
Earlier we mentioned that nodes can take on either the role
of a source, relay or destination. Energy is consumed in the
respective roles, but we are interested to find out the portion
of energy that is spent relaying. Ideally, nodes wish to spend
as little energy relaying as possible so that more energy may
be available to carry out activities that benefit them directly,
such as sending and receiving packets. In this respect, the
GTFT cooperative algorithm is seen to outperform the other
schemes. Referring to Fig 3, the portion of total energy that
nodes spent as a relay under GTFT as a network was the
least. Under GTFT, 26.3% of the energy spent as a network
was in relaying packets for others. This is an improvement
from the 27.1% achieved when Original AODV was employed. The Nuglet algorithm achieved 34.8%, which is
worse than if no cooperative algorithm was used (i.e when
Original AODV was used). However, we find that under the
Nuglet algorithm, each of the nodes spent the same portion
of energy relaying (≈33%), unlike the other schemes. This
again suggests that the Nuglet algorithm distributes the burden of forwarding packets equally between the nodes.
6.1.3 Average Energy Cost per Packet Sent
The amount of energy needed to send a data packet under
GTFT is similar to that for Original AODV. On the other
hand, the amount of energy required by the Nuglet scheme
to send out the same data packet is considerably larger. The
reason for this is that packet drops occur frequently at the
relay nodes using the Nuglet scheme. Under GTFT, a session request instead of a packet request is sent to the relay
node at the start of a session. If the request has been denied, no data packets are sent out to the relay nodes during
the duration of the session. The command center is then
notified so that configuration for the next session is started.
This way, unlike the Nuglet scheme, energy is not wasted
when data packets are sent to the relay nodes and dropped
subsequently. A session request message has a smaller payload than the data packet itself. It is therefore assumed
to be negligible and not included in the energy measurement. This explains why the energy cost of sending out a
data packet for both GTFT and Original AODV is similar.
Hence, GTFT has an advantage over the Nuglet algorithm
in that the amount of energy needed to send a packet of data
is relatively unchanged from Original AODV, and less than
when the Nuglet algorithm was employed.
6.1.4 Number of Packets Sent During Lifetime
Ideally, nodes should have a lifetime long enough for it to
send out all its data packets. As a network, this comes up
to 100,000 data packets. However, given the limited battery
power, this is not possible.
Using Original AODV, a total of 85,443 packets were
sent out before the battery lifetime expired. Meanwhile under GTFT, 85,000 data packets were sent. The slight difference is attributed to the type of sessions that were encountered by both algorithms before their battery ran out.
Under GTFT, by denying session requests occasionally, energy is conserved, and more test cases can be run. Thus,
nodes under GTFT could possibly accept session requests
in a later test case, that nodes under Original AODV would
not have heard of. The nodes in these later test cases could
be arranged in a different configuration from the one experienced by Original AODV. For example, nodes under Original AODV might have accepted 10 sessions with the laptop
as the relay and 9 sessions with the iPAQs as the relay, while
nodes under GTFT might have accepted 9 sessions with the
laptop as the relay and 10 sessions with the iPAQs as the
relay. Although the number of sessions accepted by both
algorithms are the same, the energy consumed in accepting
the sessions differs. For instance, the energy consumed in
sending out packets from the laptop is more than that from
the iPAQ. As a result of this disparity, the number of packets that nodes are able to send out during their lifetime will
differ as well.
Next, we consider the Nuglet scheme with a nuglet
counter value of 1000. Under this scheme, only 18,008
packets were sent out. This corresponds to about 21% of the
packets sent out when GTFT or Original AODV was used.
Under the Nuglet algorithm, once nodes have acquired the
nuglets necessary to send out the maximum amount of packets, packet forwarding ceases. Consequently, in the later
test cases, although all the nodes have acquired the necessary nuglets to send out the remaining packets, none of
the nodes are able to send out anything. This is because
nodes are unable to find intermediate nodes that are willing
to forward their packets to the destination node. Any future requests from the source nodes will be denied. Nodes
wait expectantly, hoping to trade in the nuglet counters that
they have acquired but are unsuccessful. Energy will then
be wasted transmitting packets only to have them dropped
at the relays.
Figure 4. Total Packets Sent as a Network
against Generosity Value
Figure 5. Convergence of NAR value
6.2.1 Increase in Packets Sent as a Network
Here, we look at how the number of packets sent by the individual nodes is affected by the generosity value. Referring
to Fig. 4, we see that by increasing the generosity value, the
total amount of packets that were sent increased considerably. This is as expected, as the more generous nodes are,
the more sessions they are willing to forward, leading to an
increase in the number of packets that are sent.
However, there is a limit to how much the nodes will
send out as a network. When the generosity value was increased from ε=0.03 to ε=0.10, we found that the energy
limitations were exceeded when the generosity value was
set to be greater than 0.05. In addition, when energy limitations were disregarded and the generosity value was increased from 0.07 to 0.10, there was no apparent increase in
the total number of packets that were sent out. This suggests
that there is no incentive for nodes to be extra generous.
6.2.2 Convergence of Normalised Acceptance Rate
6.2. GTFT Study
Besides comparing the cooperative algorithms, we also
seek to verify the simulation results of the GTFT algorithm
in [1] against an actual AHN.
The simulation results in [1] shows that if all the nodes in
the AHN employ GTFT, the Normalised Acceptance Rate
(NAR) values will converge to the Pareto optimal value.
NAR is defined to be the ratio between the number of successful relay requests and the number of relay requests that
Table 3. Summary of Algorithm Comparison
Original Nuglet GTFT
Less energy spent as a relay
Low energy cost per packet
Large no. of packets sent during lifetime X
Equitable forwarding burden
x Performs less satisfactorily compared to the best scheme
X Performs the best compared to the other schemes
have been made by the node. The Pareto optimal value is
the ideal NAR given the current network configuration. The
results obtained here shows the theoretical convergence to
be true (refer to Fig 5). In addition, we find that when the
generosity value is increased, the NAR converges faster towards the Pareto optimal point. However, when nodes reach
the optimal point they will be less willing to accept session
requests. If a node has not attained the optimal value at this
point, it may never be able to do so. This is because other
nodes seeing their session requests rejected will reciprocate
by refusing session requests. On the other hand, if the generosity value is too low, the NAR is found to be far from the
Pareto optimal point although it still converges towards it.
This could be because its convergence rate is so slow that
it is not able to converge to the Pareto optimal value within
the time frame of the experiment.
7. Conclusion
In summary, the theoretical framework and simulation
results of cooperative algorithms suggested in theoretical
papers were compared and verified against an actual AHN.
A feasibility study was also conducted to compare the merits and drawbacks of each scheme. From Table 3, we see
that GTFT gives the best performance. It reduces the energy that nodes spend relaying and allows them to send out
a large number of packets during their lifetime. While the
Nuglet scheme is not as favorable according to most of the
performance indicators here, it is able to distribute the forwarding burden among nodes for experiments conducted
in the short run. The current work assumed a single relay setup, a single session per test case and that nodes do
not misbehave. Our future work would involve observing
how the AHN performs after extending the current setup to
a multiple relay network and allowing multiple concurrent
sessions. Finally, we seek to study the effects of incorporating the measures in [6], [7] and [8] when there are misbehaving nodes.
[1] V. Srinivasan, P. Nuggehalli, C.F. Chiasserini and R.R. Rao, ”Cooperation in Wireless Ad Hoc Networks”, Proc. of IEEE INFOCOM
2003, San Francisco, April 2003
[2] L. Buttyan and J.P. Hubaux, ”Stimulating Cooperation in SelfOrganizing Mobile Ad Hoc Network”, Technical Report No.
DSC/2001/046, August 2001
[3] L. Blazevic, L. Buttyan, S. Capkun, S. Giordano, J.P. Hubaux, and
J.Y. Le Boudec, ”Self-Organization in Mobile Ad-hoc Networks: the
Approach of Terminodes”, IEEE Communications Magazine, Vol.39
No.6, June 2001
[4] M. Jakobsson, J.P. Hubaux and L. Buttyan, ”A Micropayment
Scheme Encouraging Collaboration in Multi-Hop Cellular Networks”, Proc. of Financial Crypto 2003, La Guadeloupe, January
[5] L. Buttyan and J.P. Hubaux, ”Enforcing Service Availability in Mobile Ad-Hoc WANs”, Proc. of IEEE/ACM Workshop on Mobile Adhoc networking and Computing (MobiHOC), Boston, USA, August
[6] S. Marti, T.J. Giuli, K. Lai and M. Baker, ”Mitigating Routing Misbehviour in Mobile Ad-hoc networks”, Proc. Of Mobicom 2000,
Boston, USA, August 2000
[7] S. Buchegger and J.-Y. Le Boudec, ”Performance Analysis of the
CONFIDANT Protocol (Cooperation of Nodes: Fairness in Dynamic
Ad-hoc NeTworks)”, Proc. Of Symposium on Mobile Ad Hoc Networking and Computing (MobiHOC), Lausanne, Switzerland, June
[8] P. Michiardi and R. Molva, ”CORE: A Collaborative Reputation
Mechanism to enforce node cooperation in Mobile Ad hoc Networks”, Communication and Multimedia Security 2002, Portoroz,
Slovenia, September, 2002
[9] L.M. Feeney and M. Nilsson, ”Investigating The Energy Consumption of a Wireless Network Interface in an Ad Hoc Networking Environment”, Proc. of INFOCOM 2001, Anchorage, Alaska, USA,
August 2000
[10] B. Wang and S. Singh, ”Computational Energy Cost Of TCP”, Proc.
of INFOCOM 2004, Hong Kong, March 2004
[11] A. Jain, A. Pruthi, R.C. Thakur and M.P.S. Bhatia, ”TCP Analysis
Over Wireless Mobile Ad Hoc Networks”, Proc. of 2002 IEEE International Conference on Personal Wireless Communicatons , New
Delhi, India, December 2002
[12] National Institute of Standards and Technology, ”Kernel AODV”, kernel
A Path Density Protocol for MANETs
Evgeny Osipov and Christian Tschudin
University of Basel
Computer Science Department
Bernoullistrasse 16, CH–4056 Basel, Switzerland
Email: {evgeny.osipov | christian.tschudin}
Knowing or being able to measure the “path density”
at sources of communications is essential to provide fair
capacity distribution between sessions in multi-hop ad hoc
networks. We propose and have implemented PDP, a path
density recording protocol. In this paper we describe PDP,
discuss protocol design options and explain how it was piggybacked onto an existing reactive routing scheme. We assess the validity of our approach both using simulations and
real world measurements.
Unfairness index
# of flows
2 1
# of hops
A. Plain TCP over IEEE 802.11
Unfairness index
1 Introduction
In [1] we presented an adaptive distributed capacity allocation scheme for multi-hop wireless networks. We showed
that by throttling the output rate at ingress nodes we achieve
both an increase in total network throughput and almost perfect fairness. Figure 1, generated from simulations, shows
the degree of improvement with respect to unfairness between multiple TCP sessions in static networks with different maximum number of hops. When our rate bound is
implemented we observe that the unfairness virtually vanishes.
In our solution [1] the throttling level is a function of: (a)
the number of hops for a particular “connection” (routing
path between two peers) and (b) the number of competing
connections on the path (path density). In the case that a
reactive ad-hoc routing protocol is used to provide connectivity, the first parameter of interest is easily obtainable from
the route reply (RREP) message received by the source. In
this paper we show how the path density parameter can be
obtained at run-time during the route establishment phase.
The concept of a path density (PD) implicitly exists in
the major quality of service (QoS) architectures developed
for the Internet. For instance, in services which require a
before hand reservation of network resources, this information can be extracted from the state (either aggregated or
# of flows
2 1
# of hops
B. TCP over IEEE 802.11 with our ingress throttling scheme
(see Appendix A)
Figure 1. TCP unfairness index (simulations).
See Appendix B for the definition of “unfairness index” and see Figure 5 for the topologies.
per-flow) established by a resource reservation protocol like
RSVP. In the simplest case we can count the number of entries identifying the ongoing flows in every router on the
path of a particular session and report this number to the
source. In MANETs, however, obtaining the path density
information is difficult due to frequent topology changes
and the specifics of the transmission medium. In MANETs,
the competing connections might not share common nodes
but still do compete with other connections in the carrier
sensing range.
1.1 Contribution and structure of the paper
We developed two variants of a distributed path density
protocol: A stateless PDP and a state-full PDP. To the best
of our knowledge, PDP is the only scheme which allows
the on-demand gathering of an ad-hoc network state and
which was assessed in simulations as well as real world experiments. The design of PDP greatly benefited from being
able to instantly switch between simulations and real world
experiments during the developing, debugging and performance measuring phases.
The rest of the paper is organized as follows. In Section
2 we introduce the problem of path density gathering and
describe major design options for PDP. In Section 2.5 we
show how PDP was piggybacked onto an existing routing
protocol. After that in Section 3 we report on performance
results of PDP obtained using simulations and real-world
measurements. We discuss open issues and related work in
Section 4 before concluding with Section 5.
2 Measuring the path density
Throughout the paper when talking about a stream of
packets between applications at a source and a destination
node we will use the terms connection, session or flow interchangeably.
2.1 Problem statement
The problem addressed in this paper is formulated as to
discover the number of connections competing for the transmission medium along a path of a particular session. The
problem is illustrated in Figure 2. In the figure, Connection 1 is our sample connection which attempts to discover
the path density along its path. The correct scheme should
report five cross-connections plus the main Connection 1
itself. When any two flows partly share their forwarding
paths (as is the case with flows one and four in Figure 2)
the common nodes should treat them as different connections. However when two or more connections completely
share the path from a source to a destination the forwarding nodes should treat this case as one end-to-end session.
In the later case the source nodes should perform additional
shaping actions as described in [1]. The forwarding nodes
of Connection 1 should also take care of distributing the
known information about ongoing connections in the corresponding one-hop region. In our example this is needed
to introduce the presence of flows two and three to five and
six and vice versa, since their forwarding nodes in general
might be outside the communication regions of each other.
Connection 2
Connection 5
Connection 1
Connection 6
One hop neighborhood
of connection 1
Figure 2. Counting competing connections.
2.2 General solution scheme
We developed two different approaches for gathering the
path density information: A stateless route request (RREQ)
driven and a state-full route reply (RREP) driven scheme.
Further on we will refer to the first scheme as SL-PDP
(StateLess PDP) and to the second scheme as SF-PDP
(StateFull PDP). In both approaches every node in the network maintains one state variable ND (the neighborhood
density). This is the number of cross-connections in the
neighborhood of two hops. With every new connection appearing in the neighborhood this value is incremented by
one. The state is periodically aged and is decremented by
the number of connections which were silent during this period.
The stateless approach does not keep a per-connection
state in the forwarding nodes and utilizes the fact that a
route request message indicates the desire of a node to
communicate with another node. In this scheme every
RREQ issued by the source advertises the presence of the
connection to all nodes which receive its original or rebroadcasted copy. There are two major weaknesses of this
scheme that we discovered when experimentally assessing
the functional correctness. Firstly, SL-PDP also counts the
failed connection setups as existing connections and maintains this information at least for the duration of the aging
timer. Secondly, since flooding is used for dissemination of
RREQs we cannot control the range of their distribution. As
a result, the information about the presence of a connection
is spread over more than two hops. We do not discuss further the functionality of SL-PDP in this paper and refer to
[2] for more information.
The statefull approach is free from these weaknesses
since it counts only active connections and controls the
range of information dissemination. The main idea of this
scheme is that on reception of a route reply message all involved nodes establish a state corresponding to the end-toend session. A node delays the announcement of a connection to the neighborhood until it sees the first data packet
from this connection, as only this event indicates that the
connection is active. The announcement of active connec-
tions in one hop neighborhood is done every time a node
sends or re-broadcasts a route request message. We now
describe SF-PDP in more details.
becomes active only when the first data packet arrives to
the node.
2.3.3 Adding “non-local active” SD entry
2.3 Statefull path density gathering
In SF-PDP we identify presence of a communication
session between two nodes in the network by a sourcedestination pair (SD). In this scheme every node maintains
a table of known SD pairs (SD table). The entries in this
table can be of four types: (a) Local active, (b) local inactive, (c) one-hop active and (d) two-hops active as explained
in the following subsections. The ND state variable is the
number of all “local active” and “non-local active” entries
in the SD table. The major operation of the statefull PDP is
shown in Figures 3 and 4.
The non-local active entries are those which identify connections ongoing in the neighborhood of two hops but not
passing through this node. The information about such connections is distributed with broadcast route request messages for any connection originated or re-broadcasted from
any of the one-hop neighbors. Figure 4 illustrates the principle of such information dissemination. Before issuing or
re-broadcasting the route request message, SF-PDP examines the local SD table and appends all “local active” and
“one-hop active” SD pairs to the RREQ SD List. After that
it piggybacks the constructed list to the RREQ messages
and transmits it.
//Field is_local: 1 (local active),
−1 (local inactive)
0 (one−hop active), 2 (two−hops active)
2.3.1 Adding “local inactive” SD entry
The operation of the statefull PDP during registration of a
connection in the network is shown in Figure 3. When a
route request message for a particular connection successfully reaches the destination the RREP message is returned
to the source. However, reception of the RREP message by
a node does not indicate the success of the route establishment procedure since the message itself can be lost on the
path to the source. SF-PDP treats the event of RREP reception as an indication of a possible connection through this
node. Upon reception of a RREP a SD entry corresponding
to this connection is inserted to the SD table and is marked
as inactive.
ND is unchanged;
ins_SD(S,D,is_local=−1); //Local inactive
Source (S)
DATA pkt1
DATA pkt1
DATA pkt1
Destination (D)
DATA pkt1
ins_SD(S,D,is_local=1); //Local active
Figure 3. SF-PDP: Processing of RREP and a
first packet.
IF (is_local==1) THEN write((S,D,0),RREQ_SD_List);
IF (is_local==0) THEN write((S,D,2),RREQ_SD_List);}
Source (S)
Destination (D)
//Format of SD record: {S,D,location}
//location: 0 (one−hop) or 2 (two hops)
IF (!known_SD) THEN ins_SD(S,D, is_local=location);
IF (known_SD) THEN update_SD(S,D,is_local=location);
Figure 4. SF-PDP: Processing of RREQs.
When a route request message is received by a node,
the SD entries from the RREQ SD List are inserted to the
local SD table and marked as “non-local active” of the
corresponding location. After that SF-PDP removes the
RREQ SD List from the received RREQ message and appends its own list as described above. By replacing the
RREQ SD List with every retransmission we do not allow
the information about local connections to spread further
than one hop.
2.3.4 Aging of SD entries
Every entry in SD Table is assigned a timer. The entry is
deleted when the timer expires. The duration of the timer is
the same as of the route expiration timer.
2.3.2 Adding “local active” SD entry
The only definite indication of the active connection in any
forwarding node is the event of reception of the first data
packet. Therefore, when a local inactive SD record is created, the forwarding engine is instructed to catch the first
packet in either direction of the connection.The SD entry
2.4 End-to-end density reporting and smoothing
The overall goal of PDP is to deliver the path density
information to the source nodes. This information is used
by the interface queue to configure the delay of emission of
locally generated data packets as described in [1].
The path density is the maximum value of ND variables
from all forwarding nodes on the path of a connection. The
ND value is piggybacked in the route reply message at the
destination node and can be modified by every node which
receives and forwards the RREP message. When a forwarding node receives a route reply message, PDP examines the
local ND state and compares its value to the ND field in the
RREP message. The higher value proceeds further to the
As we see the path density state is refreshed with every
route reply message. Therefore the frequency of its update
depends on the frequency of route refreshes initiated by the
source. Here we assume that route refreshes are periodically initiated. In the next subsection we discuss how this
requirement can be combined with reactive ad-hoc routing
Due to high dynamics with which new flows might enter and leave the ad-hoc network, the path density reported
to the source may fluctuate in time. Therefore, in order to
avoid reconfiguring the scheduler on the interface queue too
often we allow sources to smooth the path density by computing a sliding average.
2.5 Integration in ad-hoc routing protocols
Our primary goal for the implementation of PDP was to
avoid introducing additional message exchanges to gather
the path density information. This is because from the functional point of view the operations of PDP are very similar
to operations of the route establishment phase. Implementing an additional stand-alone protocol for dissemination of
PDP information would in the worst case double the capacity already consumed by a routing protocol. We decided to
re-use the existing mechanism of message dissemination of
the ad-hoc routing schemes.
We integrated PDP in LUNAR [6], which is a L2.5 reactive routing scheme. We also considered possibility of PDP
integration in other reactive routing protocols, i.e AODV
[4] and DSR [5], and in the proactive OLSR protocol [3].
In this subsection we present our observations on the possibility to use AODV, DSR and OLSR protocols as a PDP
The route refresh period in proactive routing is normally
larger than in reactive schemes. In OLSR, for example, the
default refresh period is 16 seconds. We consider it too large
assuming high dynamics with which new sessions might enter the ad-hoc network. Based on these observation we do
not see proactive routing protocols suitable for dissemination of PDP information.
As for AODV and DSR, we see a number of difficulties
for possible integration of PDP in these protocols. As stated
in Section 2.1, the flows which share a part of their path to
the destination should be treated by our scheme as separate
connections. This requirement places a major limitation on
the carrier protocol: It should not perform path aggregation.
However, this functionality is embedded in both AODV and
DSR. Firstly, path aggregation is the normal operation for
routes to the same IP subnet. Secondly, even for the cases
where IP level aggregation is not used, AODV and DSR
provide gratuitous route replies: When a forwarding node is
part of the path to a particular destination, it may return the
RREP to the source without further propagation of the route
request message. Integration of PDP into AODV or DSR
would require changes to the core functionality of these L3
routing protocols.
2.6 Integration of PDP in LUNAR
LUNAR is a L2.5 reactive routing scheme. The major
difference of LUNAR from AODV and DSR is its position
in the TCP/IP protocol stack. Since LUNAR is located below IP layer, the IP level route aggregation which is present
in L3 routing schemes is impossible. The route request messages in LUNAR always propagate from the source to the
destination. That is, gratuitous replies are not used either.
LUNAR is well suited for integration with PDP also because of its high dynamics of information update: LUNAR
re-establishes the entire path every three seconds.
3 Experiments
The single code base both for the Linux kernel and the
network simulator ns-2 [12] allowed us to debug and extensively test the PDP+LUNAR functionality on a wide range
of static and dynamic simulation scenarios. We implemented and tested both stateless and statefull PDP schemes.
Once we obtained a stable protocol we generated a real
world version and performed a correctness test in a test-bed
as described in Appendix C. While performing a preliminary evaluation of SL-PDP in the simulator we found the
drawbacks of this scheme described in Section 2.2. As a
result we decided to go on with SF-PDP. In this section we
present the results from testing the protocol both in simulations and the real-world test-bed.
3.1 Metrics
In all experiments we used correctness and convergence
time metrics to evaluate the performance of PDP. Correctness is evaluated with respect to the “path density” as reported to the end nodes of the session (the path density in
our experiments was also reported to destinations because
we used bidirectional traffic in the experiments).
First, we show the effect of path density on the unfairness index and total TCP throughput in networks with different number of active TCP flows. We performed a series
of simulations in ns-2 (see Appendix C for setup details) on
a set of three hop networks depicted in Figure 5. We varied
the number of connections in the network from 2 to 9; For
each case we run 30 simulations with enabled and disabled
reaction on the reported path density. The dynamics of the
unfairness index (see (2) in Appendix B) in both cases is
shown in the part of surfaces corresponding to three hops
and 2 – 9 connections in Figure 1.
Under our scheme the fairness among competing TCP
flows was close to perfect, while ignoring the path density
we observed an unfairness of up to 25%. Note that the nature of the used unfairness index reflects the closeness of
individual goodputs of each flow to the ideal share, therefore the closer the unfairness index to zero the higher is the
minimal individual goodput.
Number of
Total TCP throughput, kb/s
Without path density With path density
Table 1. The effect of path density on total
TCP throughput.
Table 1 shows the effect of ingress rate throttling on total
TCP throughput depending on the path density for the two
cases. As we observe from the table the total TCP throughput in the network is higher when accounting for the path
density. This implies better network utilization when all
competing flows fully utilize their allocated share. Note that
in reality we will also have a number of short-living communications and flows with a number of short transmission
bursts e.g. web browsing. In this case the share of capacity
computed based on the path density reported by PDP will
not be fully utilized. When several connections are originated from the same source the leftover capacity can be
reused by the active flows, otherwise the network will be
underutilized until the route for the inactive connection will
be removed. Therefore it is essential to dynamically update the ingress nodes with the path density information. In
the case of our PDP+LUNAR combination the information
about inactive flows will be aged within three seconds in all
PDP-aware forwarding nodes.
TCP src 1
TCP dst 1
TCP dst 2
TCP dst N
250 m
3.2 The effect of path density (Simulation)
TCP src 2
TCP src N
126 m
Figure 5. A set of three hop networks for illustration of the path density effect.
3.3 Static Scenarios (Real World)
We present an evaluation of the functional correctness
of PDP on two static scenarios depicted in Figure 6. While
all nodes in reality are located in the reception range of each
other (hence the same radio interference domain) we configured LUNAR so that a node can hear the data transmission
only from a selected set of nodes and discard packets from
others. In this sense only Nodes 3 and 4 in Figure 6a “share”
the regions of assured data reception of each other. In the
second case (Figure 6b) Node 2 is able to hear transmissions from Node 4 and Node 3 from Node 5. PDP should,
however, report the same density information on all nodes.
In both scenarios Session 1 started at time 1 s and finishes at time 30 seconds. Session 2 establishes the path 10
seconds after Session 1. The duration of Session 2 is 10
seconds. We used the ping protocol to generate traffic of
the corresponding sessions. The interval between two consequent ICMP requests is 100 milliseconds. We repeated
both real world experiments up to 10 times and obtained
the same behavior of our two metrics. Figure 7 shows the
results from one of these runs.
As is visible from the time diagrams, PDP correctly reports the path density to the corresponding end nodes of
the two connections within 1.5 – 3 seconds from the start
time of a session. The delay in dissemination of the information is explained by the default route refresh interval in
LUNAR. After the session is established the earliest time a
new RREQ message is generated by the destination in the
backwards direction is after 1.5 seconds.
3.4 Dynamic Scenario (Simulation)
We evaluated the behavior of PDP in presence of mobility in simulations only. This time we seek a quantitative assessment of the path density information provided by PDP.
We studied the scenario shown in Figure 8.
Session 2
Node 1
Node 2
Session 1
Node 3
Node 1
Session 1
Node 3
Session 2
a) Topology 1.
Node 4
Node 5
Node 6
560 m
Node 5
500 m
Node 4
240 m
Node 2
Node 8
Session 2
Node 1
Session 3
Node 7
Node 2
Node 3
Session 1
Node 4
Figure 8. Topology 3 for experiments on a dynamic scenario.
Node 5
b) Topology 2.
Figure 6. Topologies
Session 1 Session 2
Session 2
Session 1
Session 1 Session 2
Session 2
path density
Path density at source nodes
Path density at source nodes
Reported path density
Node 4
Node 4
Node 3
Node 2
Node 1
Time, s
a) For scenario in Figure 6a.
Node 2
Node 3
Session 1
Node 5
Node 5
Node 1
Node 9
Time, s
b) For scenario in Figure 6b.
Figure 7. Reported path density in real-world
quence numbers curves is the same for all three flows provided path density is taken into account, which results in fair
sharing of the network’s capacity. Moreover, in this case
the progress of all TCP flows is smooth and free from interruptions when compared to the case where the path density
information is ignored. In addition to the smoothness of the
TCP flows, the resulting total TCP throughput (not shown
in the figure) remains the same as in the case without rate
limitation at sources.
To sum up, the results of the last experiments illustrate
our overall goal: If flows in a MANET are aware about the
presence of each other and reduce their rate accordingly,
none of the competitors is able to capture the capacity as
it happens for the plain combination of MAC 802.11 and
TCP. The latter case is represented in Figure 9 by flow TCP
2 (w/o path density) which worsens the performance of TCP
flows 3 and especially 1.
4 Discussion and related work
In the scenario we have three TCP sessions each following a path of two hops. Initially all nodes are located outside
the range of assured reception of each other. After five seconds from the simulation start the middle nodes of sessions
one and three begin a movement towards the middle node
of Session 2. The speed of the nodes is 10 m/s. In their final
destination both Nodes 2 and 8 are in the range of assured
reception of each other and Node 5. The mobile nodes remain in this position for 80 seconds; After that they start
moving back to their original position.
The goal of this experiment is to evaluate the dynamics
of PDP based source throttling. We performed two experiments on this scenario. First, the path density information
was ignored and TCP flows might transmit without rate limitation. In the second experiment we configured the scheduler on the interface queue with a throttling bound dynamically computed according to the path density information
reported by PDP. Figure 9 presents the results of these experiments.
As we observe from the figure, the slope of TCP se-
Our path density gathering scheme presented in this paper is a distributed protocol for the on-demand discovery of
an ad-hoc network state. The PDP protocol plus our ingress
rate throttling approach permits to implement a capacity allocation scheme with guarantees on fairness of communications.
In addition to the admission control in sources we see
potentials of our protocol for congestion-aware routing. Indeed, the ND state kept in all forwarding nodes is an excellent indicator of the neighborhood’s load. If a forwarding
node – instead of re-broadcasting the newly arriving route
request – would first examine the neighborhood density, it
can estimate the chance for this flow to obtain an acceptable
service while being forwarded through this area. If a neighborhood is overloaded with existing connections the route
request can simply be discarded. Assuming a network with
relatively high node density, the “surviving” route requests
would discover less loaded paths.
Path density Node 7
Flow 3
w/o path density
w path density
Path density Node 4
5 Conclusions
Flow 2
w/o path density
w path density
Flow 1
Path density Node 1
such information is however not described in the paper. The
authors suggest to use additional message exchanges for the
dissemination of the information.
In [9] the authors discuss a distributed weighted fair
queuing algorithm, similar to its analog in the fixed Internet.
In order to distribute the weights to the interface queue of
each node in the neighborhood of one hop, the authors suggest to encode this information in the MAC 802.11 header.
w path density
w/o path density
Time, s
In this paper we considered the problem of measuring the
“path density” in MANETs. We presented an approach for
gathering such information in a distributed way at run-time.
We understood that building PDP based on observing route
request transmissions only leads to incorrect counting of the
failed connections and resorted to storing per-connection
state in the forwarding nodes. We showed how PDP can
be piggybacked inside LUNAR messages without introducing new transmission events and tested our implementation
with combined simulation and real-world measurements.
Figure 9. Path density and TCP sequence
numbers progress for scenario in Figure 8.
4.1 Related work
The idea of estimating the state of an ad-hoc network is
not new. There are several approaches based on distributed
network state exchange. In [7] SWAN, a distributed QoS architecture for MANETs is proposed. The proposed admission control at sources as well as the rate limitation scheme
in all nodes is based on hop-by-hop measurements of the
available bandwidth. The authors suggest to use a reactive
routing protocol to gather the results of these measurements
along a path of a connection. New flows are admitted for
transmission only if the reported path capacity is less than a
certain threshold.
In [8] the authors suggest to exchange the occupancy
level of the interface queue of every node in one and two
hops neighborhood. This information is then used to construct a virtual queue of the neighborhood and to coordinate dropping of packets. The exact mechanism to exchange
We wish to thank an anonymous reviewer for the detailed comments which greatly helped the preparation of the
camera-ready version of this paper.
[1] E. Osipov, C. Jelger, Ch.Tschudin, “TCP capture
avoidance in wireless networks based on path length
and path density”, Technical Report CS-2005-003,
University of Basel, Switzerland, April 2005.
[2] E. Osipov and Ch. Tschudin “A path density protocol
for MANETs”, Technical Report CS-2005-004, University of Basel, Switzerland, May 2005.
[3] T. Clausen, P. Jacquet, “Optimized link state routing
protocol (OLSR)”, RFC 3626, IETF, Oct 2003.
[4] C. Perkins, E. Belding-Royer and S. Das, “Ad hoc ondemand distance vector (AODV) routing”, RFC 3561,
IETF, Jul 2003.
[5] D.B.Johnson, D.A.Maltz,Y-C.Hu,
“The dynamic
source routing protocol for mobile ad hoc networks
(DSR)”, IETF draft (work in progress), 2003. [Online]. Available:
[6] Ch. Tschudin, R. Gold, O. Rensfelt and O. Wibling, “LUNAR: a lightweight underlay network adhoc routing protocol and implementation”, In Proc.
NEW2AN’04, St. Petersburg, Russia, Feb. 2004.
ringress =
α · Lcrit
[7] G-S. Ahn, A. T. Campbell, A. Veres, L.
IETF draft (work in
progress), 2003. [Online]. Available:
In (1) Lcrit
T CP is the estimated rate of TCP data segments
which creates the critical network load, h is the number of
hops of a particular flow and N is the path density. If all
sources follow this rule, then the total network capacity is
fairly distributed between the competing flows as shown in
Figure 1.
[8] K. Xu, M. Gerla, L. Qi, and Y. Shu, “Enhancing TCP
fairness in ad hoc wireless networks using neighborhood RED” , In Proc. MobiHoc’03, Annapolis, MD,
USA, 2003.
Appendix B - TCP unfairness index used in
Figure 1
[9] N.H. Vaidya, P. Bahl and S. Gupta, “Distributed
fair scheduling in a wireless LAN”, In Proc. MobiCom’00,Boston, MA, USA 1999.
We define the unfairness index u used in Figure 1 as the
normalized distance (2) of the actual throughput of each
flow from the corresponding optimal value.
[10] K. Xu, S. Bae, S. Lee and M. Gerla, “TCP behavior
across multi-hop wireless networks and the wired Internet” , In Proc. WoWMoM’02, Atlanta, GA, USA,
Sep. 2002.
[11] Uppsala University Ad Hoc Implementation Portal,
[Online]. Available:
[12] Network Simulator NS-2,
[Online]. Available:
Appendix A - Adaptive capacity allocation
scheme for capture-free communications
TCP capture, as is described in [10], is one of the unsolved problems in multi-hop ad hoc networks which results in extremely unfair distribution of network bandwidth
between competing sessions.
In [1] we suggested a scheme for dynamic capacity
distribution between competing connections that partly or
completely share communication regions. We proposed the
ingress throttling mechanism which guarantees capture-free
communications for all involved flows.
The input information to our scheme is the number of
competing connections on the path of a particular multi-hop
TCP connection (path density) and the knowledge of the
saturation point of the radio transmission medium beyond
which a single TCP flow starts to loose packets (critical network load). This information is used to compute the output
rate at ingress nodes (1).
(Xopt − Xacti )2
qP i
In this formula Xopti is the ideal throughput of flow i
obtained under fair share of the network capacity. In order to compute this value we divide the throughput of the
corresponding flow obtained when running alone in the network by the number of competing flows. Xacti is the actual
throughput of the same flow achieved while competing with
other flows. This index takes the values between 0 and 1
and reflects the degree of global user dissatisfaction, hence
the value of 0 corresponds to perfectly fair communications
and 1 represents the opposite case. Note that the index is
useful only when applied to two or more flows. This is also
reflected in Figure 1, where the y-axis starts from two connections.
Appendix C - Experiment Setups
The real-world results were obtained from a setup of five
DELL Latitude laptops with ZyXEL ZyAIR B-100 wireless interfaces. We use the Linux operating system with 2.6
kernel and LUNAR implementation available from [11].
The simulation results were obtained using ns simulator
of version 2.27. In simulations with scenarios depicted in
Figures 5 and 8 we used TCP Newreno as the most popular
variant of the protocol. We set the value of the TCP maximum segment size (MSS) to 600 B. The data transmission
rate of all devices is 2 Mb/s. Other ns-2.27 parameters have
default values.
Interactions between TCP, UDP and Routing Protocols
in Wireless Multi-hop Ad hoc Networks
Christian Rohner, Erik Nordström, Per Gunningberg, Christian Tschudinb
Uppsala University, Sweden and University of Basel, Switzerlandb
{chrohner|erikn|perg}, [email protected]
The achieved throughput in an ad hoc network is affected by many factors, including radio interference between hops, the ability of the routing protocol to react on topology changes and a complex interaction between the application and underlying protocols. This
paper studies experimentally the impact of these factors on UDP and TCP throughput. Furthermore, when
both TCP and UDP share some hops in an ad hoc network there is a complex interaction between the transport protocols as well as with the routing protocol. Our
results show that this interaction results in significant
UDP jitter, instable routes and significantly lower TCP
throughput. We use a controlled real testbed for the experiments and a graphical tool that captures the interactions.
It is well known that TCP suffers in a wireless environment because it incorrectly interprets packet loss as
congestion. It is also well known that both UDP and
TCP performance is reduced in a multi-hop 802.11 environment because of radio interference between hops.
Simulation results by Holland et al. [5], show that
the radio interference for two hops reduces the possible throughput to 50 percentage. For three hops the
throughput is 33 percentage. This result was supported
both by simulations and a theoretical model.
Another performance impairing factor is frequent
routing updates. In ad hoc networks with mobile nodes,
the problem for TCP is getting more severe since it is
likely that packets get lost when a route is lost and the
routing protocol needs to discover a new route.
Finally, it is also known that UDP is unfair to TCP
when they share a common bottleneck link. TCP will
back off and try to find a sending rate that adjusts to the
rate of UDP. This adjusting time can be long depending on end-to-end delay and assumes some stability in
available bandwidth. All these factors impact the performance of an ad hoc network and have been individually
studied, but the interactions between them in a dynamic
environment is not generally understood.
The intention with this paper is to demonstrate measurements results and to study unforeseen interactions
between components at different layers that negatively
impact the experienced application level performance
in the ad hoc network. With our tools we can capture the series of transmissions involved and explain the
timing of events relative to the node movements. We
do this in a controlled experimental environment with
the possibility to repeat experiments to understand the
variance. The scenarios are carefully chosen to reflect
the intended interaction between protocols, radio range,
movement and prospective applications.
All scenarios are derivatives of the scenario shown in
Figure 1. It comprises three nodes that have contact only
with their adjacent neighbors. A low rate UDP flow,
traversing two hops, is sent from node 2 to node 0. This
scenario constitutes our baseline case, which remains
static and is never changed in any of the scenarios. We
then add a fourth node, sending an additional TCP flow
to node 0. We also add dynamic routing and mobility.
We do this to study the effect on the baseline case in
terms of interference, contention for the wireless channel and buffer space. The low rate UDP flow is used to
sample the wireless channel so that we can infer the interference from the fourth node. In theory, the expected
result would be that the fourth node will have a limited
effect on the UDP flow, because TCP should adapt to
the remaining bandwidth.
Figure 1. Base scenario: A static two-hop setting with a constant bit rate UDP flow.
From our measurement results we find that adding
the fourth node interferes with UDP, inducing jitter
and packet loss. Adding mobility and dynamic routing causes an average packet loss of up to 21.6% on
the UDP flow. The results indicate that TCP and UDP
contend for both buffers at nodes and radio transmission time. Both protocols suffer from significant loss as
expected, and TCP backs off. Somewhat unexpected is
that UDP losses are higher than could be expected on
shared wireless links. This raises the question if TCP
can adapt quickly enough to the environment where
there is node movement and connectivity changes.
Another observation is that the routing protocol is
also affected by the TCP flow, preventing the establishment of correct routes during extended periods. As
TCP will back off and then start to probe again for
the sustainable bandwidth, this potentially leads to long
term oscillations overlaying the short term interactions
among the protocols.
The paper is structured as follow. In section two we
discuss related work. In section 3 we describe our experimental set-up, scenarios, experiments and results. We
conclude with a section on discussion and further work.
Related Work
The TCP feedback loop is responsible for adapting
the sender data rate in response to, e.g., congestion.
However, this end-to-end congestion control mechanism has reduced efficiency in wireless networks because transmission is inherently broadcast. Furthermore, there are different and transmission rate dependent ranges for unicast transmission, broadcasts and interference [2].
Holland et al., examine in simulation the TCP performance over mobile ad hoc networks. They note that
TCP suffers significantly in mobile, multi-hop scenarios, simply because TCP can not distinguish between
link failures and congestion.
In [3], Gerla et al. study through simulation, the interactions between TCP and different MAC layers. Our
work complements that work by also looking at other
layers and different transport protocols.
Gupta et al. report in [4] on decreased TCP performance in the presence of interacting UDP flows. Their
study is done in the ns-2 simulator on an artificial grid
topology and with relatively high UDP data rate (800
Kb/s). Since UDP lacks rate adaptation and congestion control, a high data rate CBR flow will naturally
congest the channel. Our work complement that work
by performing real world measurements with UDP and
TCP data flows in scenarios with increasing complexity. Moderate data rate UDP is used (typically streaming
music) which – as we believe – is more realistic
Internet routers improve fair access to its resources
by using Random Early Detection (RED). Xu, Gerla,
Qi and Shu study in [8], TCP fairness in wireless ad
hoc networks. They look at queuing policies to improve TCP fairness and show that the RED scheme fails,
because flows compete not only for queue space but
also for the wireless channel. They propose Neighborhood RED (NRED) where all individual queues of
nodes within a neighborhood is treated as a shared distributed queue. Based on the notion of this distributed
queue, a forwarding node can drop packets to increase
TCP fairness.
This section reports on real world experiments that
examine the contention between TCP and UDP data
flows. After providing some setup details, we discuss
a sequence of increasingly more complex scenarios and
the results of the following experiments.
All experiments were conducted indoors in a systematic way using the APE testbed [7]. APE provides
a test environment that allows repeatability of experiments and provides support for extensive logging and
analysis of experiment data. The APE testbed orchestrates the scenarios and provides verbose logging on the
network interface and network layer. Standard laptops
(IBM X31) with Lucent/ORiNOCO IEEE 802.11b network interfaces are used, transmitting at a fixed rate setting of 11 Mbit/s, without RTS/CTS. The interface used
a standard MTU setting of 1500 bytes. Iperf and Ping
were used to generate data. For UDP, 1470 bytes data
was sent1 , resulting in 1512 byte frames in the ether.
This size was used to compare against the normal TCP
segment size. Logged and time stamped data is time
synchronized between nodes using periodic broadcasts
of time information at a rate of one packet every 10 seconds. This time information is used to synchronize the
experiment data collected from the different nodes.
The AODV-UU [1] implementation was used where
dynamic routing was needed. AODV-UU was chosen
because it is mature and has been interoperability tested.
Bandwidth Measurements
As a calibration and verification to prior work, we
start by measuring the maximum achievable throughput
for TCP and UDP data flows. We do this to establish the
expected bandwidth for our baseline scenario, in single
and multi-hop settings without mobility. For this experiment, four nodes are placed in a linear chain such that a
node can only communicate with its adjacent neighbors.
The UDP and TCP throughput over one to three hops
are measured. The results are shown in Table 1. We
differentiate in the one hop measurements between the
two nodes located next to each other (short hop) and located in the periphery of each others transmission range
(1 hop).
short hop
(10 cm)
1 hop
(ca. 20 m)
2 hop
3 hop
Table 1. The maximum throughput is decreasing with increasing number of hops.
The UDP throughput is as expected higher than the
TCP throughput, since UDP is one way traffic without
acknowledgments that compete for transmission time
over the shared channel. The maximum achievable TCP
throughput over one hop is roughly a third of that of
UDP. Acknowledgments going in the reverse direction
is not as large as the data packets in the forward direction, giving more channel time for the data.
When increasing the number of traversed hops, the
throughput decreases for each hop, in case of TCP following roughly an 1/n slope where n is the number of
1470 bytes is the default Iperf setting.
hops. This result matches the simulation results from
Holland, et al. [5].
A noteworthy observation is the significant difference
in one hop throughput between nodes that are placed
next to each other (short hop, 10cm), and nodes that
are located in the periphery of each others transmission
range. In our indoor environment, that was a separation
of 20 m. An analysis of the TCP experiments shows a
slightly higher variation in the round trip time for separated nodes, indicating retransmissions on the link-layer
as one reason for that artifact. TCP is more affected than
UDP because of its own adaptation mechanisms that are
not designed for fluctuating wireless environments.
Baseline Scenario: Static Multi-hop UDP
This experiment presents our baseline scenario
shown in Figure 1. It consists of three nodes 0, 1, and
2 placed in line. Each node has connectivity only with
its adjacent neighbors. End node 2 sends a constant bit
rate UDP flow to node 0, over the intermediate node 1.
As well as establishing the baseline performance, the
experiment intends to examine the interference between
the consecutive hops in the multi-hop path.
For each test run, the offered data rate (rate of data
from the application) is increased, permitting to study
the effect of interfering hops on the multi-hop path.
Since the bandwidth at node 1’s network interface is
shared between the two “links” we expect the overall
UDP throughput to decrease when the offered data rate
at node 2 exceeds one half of the bandwidth at node 1.
Figure 2 shows the received data rate as a function of
the transmitted data rate (rate of the data actually sent on
the channel). It can be observed that the transmitted data
rate for this two hop path achieves the best received data
rate at around 3.6 Mbit/s. Increasing the offered data
rate beyond that has a negative impact on the overall
throughput, because of interference between the hops.
Note that although we increased the offered data rate
up to 10 Mbit/s, the transmitted rate never exceeded 5.2
After having determined the interference limited data
rate over two hops, the UDP data rate was fixed at 192
kbit/s. This is well within the limit of link interference
and also resembles the rate of an MP3 data flow. We expect such a low rate flow to have close to 100% delivery
ratio in our baseline scenario with no interference.
The average UDP delivery ratio over five test runs
was 99.2 %. It could be verified that packet losses oc-
t_packet [s]
t_packet [s]
Figure 2. Received bit rate as a function of
the transmitted bit rate in the two-hop scenario
for offered bit rates between 0.5 Mbit/s and 10
Mbit/s (with 0.5 Mbit/s intervals in between).
The line shows the optimal throughput when
there is no packet loss.
curred at the second hop link between node 1 and 0.
The interaction between the two hops is analyzed by
looking at the UDP packet interspacing time. A variation in the packet interspacing time indicates either contention for medium access, link layer retransmissions,
or packet loss. Figure 3 shows the packet interspacing
time for the UDP flow between node 2 and 0 for a representative run. We observe that the second hop is more
affected by interactions than the first link. However, as
expected, the packet loss is low and packet interspacing small. In the following experiments we use these
baseline results to study the effects on the UDP flow by
adding an extra node, a TCP flow, dynamic routing and
Multi-hop UDP With an Interfering TCP Flow
In this scenario we extend the baseline scenario with
an interfering TCP flow. An extra node 3 starts at the far
left position in Figure 4, outside the transmission range
of node 1. Ten seconds into the scenario it starts a TCP
connection to node 0. After another ten seconds, node
3 starts moving toward node 0 where it stops its movements at time 38 but continues to send data. At this position, node 3 is within the transmission range of node 1
and hence will interfere.
Routing in this scenario is static. The purpose is to
see how TCP adapts its send rate (if at all) when mov-
time [s]
Figure 3. Example of UDP packet interspacing time in the static multi-hop scenario (192
ing to a position where it is potentially more affected by
channel contention from both node 1 and node 2, without an actual change in the path. We also want to see
how the UDP flow from node 2 to node 0 is affected
by the increased contention. We do this to better understand the effect of mobility and spatial placement of
Figure 4. Scenario for two-hop UDP with an interfering one-hop TCP flow.
Over five test runs, the average UDP delivery ratio
was 97.4 %. Although a slight decrease from the previous scenario, it does indicate that the UDP flow is not
significantly affected by the interference from the TCP
In Figure 5, we see the UDP packet interspacing time
on the link between node 2 and 1, and 1 and 0 respectively, along with the TCP time sequence graph for the
TCP flow between node 3 and 0, for one of the experiments. Although variations are apparent throughout all
experiments, Figure 5 is representative for a typical test
As expected, the variation of the packet interspacing
t_packet [s]
t_packet [s]
UDP 1>0
seq. nb. [1k]
UDP 2>1
time [s]
mobility, multi-hop TCP, and dynamic routing using
AODV-UU. The UDP flow, again, is sent from node 2 to
node 0. Node 3 now starts out alongside node 0 at position A, as illustrated in Figure 6. Node 2 starts its UDP
flow to node 1 simultaneously as node 3 starts a TCP file
transfer to node 0 and starts moving towards position D.
After 62 seconds it will reach position D and then turn
back and move towards node 0 again. During this time,
the TCP flow to node 0 is sent over a path that increases
from one hop to two and three hops, and reduced back
to one hop as node 3 is on the way back. Note that in
this scenario, TCP and UDP have competing data flows
going over the same intermediate nodes.
Figure 5. Example of UDP packet interspacing
time and TCP sequence graph.
time between node 2 and 1 is more apparent after node
3 is in direct contact with node 1. It interesting though,
that the packet interspacing time between node 1 and 0
tends to stabilize during the static periods (time 5..10,
and after time 50 where basically the variation pattern
from the first hop gets propagated).
The TCP flow itself is also affected. In the time sequence graph we observe some interruptions in the TCP
flow of up to 500 ms, in particular during the movement
of node 3 (time 20..38). Furthermore, TCP throughput
is unstable even during the stationary phases at the beginning and end of the experiment. While the maximum achieved TCP throughput (1.63 respectively 3.37
Mbit/s) matches the one-hop figures of Table 1 where
no UDP background traffic was present, the average
throughput during these phases is significantly lower
(0.89 respectively 1.66 Mbit/s).
An interesting observation can be made at time
50, when there is an immediate increase in the TCP
throughput. This behavior is persistent in most of the
experiments. It is not clear why TCP increases its
throughput in this situation, at the same time as the fluctuations on the second UDP hop seem to stabilize. One
possible explanation could be the capture effect [9]. Understanding the exact reasons for this particular behavior
requires further investigation.
Multi-hop UDP Sharing Hops with a TCP Flow
Figure 6. The Roaming Node scenario.
From separate analysis we know that there is one hop
connectivity between node 3 and node 0 up to about
time 25, followed by two hop connectivity up to time
50, and then three hop connectivity until time 92 when
the route switches back to a one hop configuration until
the end of the experiment. This is due to how AODV
works; on the way back it will not optimize the route
until a HELLO message is received by node 3 from the
node 0.
The average UDP delivery ratio in this scenario decreased to 78.4% and the average TCP throughput to
0.93 Mbit/s 2 . A summary of the different scenarios in
UDP deliver ratio and TCP throughput is shown in Table
Figure 7 shows the UDP packet interspacing time and
TCP sequence number graph for one of the experiments.
It can be observed that the TCP data flow stalls in particular on the change from a one hop to a two hop configuration (time 25..30), and on the change from a two hop
to a three hop configuration (time 40..50). There are
also sporadic stalls during the three hop configuration
(time 50..92). The different configurations are visible
in the (slightly) decreasing slope of the time sequence
graph. The switching of routes on the way back is much
This scenario, called “Roaming node”, extends the
complexity of the baseline scenario by adding more
In the Roaming node scenario we experienced an outlier in one
of our five test runs. Those results were therefore discarded and the
average calculated over the remaining four runs.
Static multi-hop UDP
Multi-hop UDP with TCP flow
Roaming node
UDP Delivery Ratio
99.2 %
97.4 %
78.4 %
TCP Throughput
1.34 Mbit/s
0.93 Mbit/s
t_packet [s]
UDP 1>0
seq. nb. [1k]
UDP 2>1
t_packet [s]
Table 2. Average UDP delivery ratio and TCP throughput for the different scenarios.
time [s]
Figure 7. Example of UDP packet interspacing
time and TCP sequence graph in the Roaming
node scenario.
smoother. This can be explained by AODV’s HELLO
messages working more proactively on the way back,
discovering a more optimal (shorter) route before the
old one is gone.
We further observe increased UDP packet interspacing on the link between node 2 and 1 during three hop
TCP connectivity (time 50..80). Increased UDP packet
interspacing on the link between node 1 and node 0 is
observed whenever we have progressing TCP traffic, but
the UDP flow immediately stabilizes during TCP stalls.
The increased packet interspacing is caused by the extra
queuing delay on the intermediate nodes when the TCP
flow also competes for buffer space.
An interesting observation can be made in the beginning of the experiment. It seems that when the TCP
and UDP flows start simultaneously and in combination
with dynamic AODV routing, the multi-hop UDP path
between node 2 and 0 is broken. We see two different
explanations for this (or likely a combination). Either
TCP captures the channel, causing significant loss on
the first hop UDP link (consequently no traffic is seen
on the second hop link either). A more probable explanation, though, is that TCP’s aggressive start impacts
AODV’s HELLO neighbor sensing and broadcast route
discovery. The probing packets used in these mechanisms are broadcasted on the link layer, hence without
acknowledgments or retransmissions. If these packets
are lost, no path will be established. Not until time 10,
after some movement will TCP back off and allow the
UDP flow between node 2 and node 0 to resume. The
reason for this is probably that node 2 was previously a
hidden terminal for node 3, causing massive collisions
at node 1. Only after node 3 moves within contention
range of node 2, will it back off. The exact cause of this
problem will be further examined in the next section,
where we present a deeper analysis of protocol and link
Analyzing Protocol and Link Interactions with
“Activity Plots”
The experimental analysis in the previous section
raises many questions about the causes of the seemingly random and complex performance anomalies. It
is necessary to study the interactions of protocols and
link conditions at a much finer granularity in time to understand what is really happening. For this purpose we
have developed an analysis tool to capture interactions
between the protocols at different layers and for different nodes in a graphical way. This tool enables us to
create “activity plots” from the collected data. Activity
plots are a way to visualize and understand the timing
of events. They complement the traditional flow based
analysis with a more detailed analysis of activity on individual links. The idea is to present an experiment’s
complete set of events at a time granularity that is between single trace entries and aggregated performance
An activity plot shows link activities on each nodes’
wireless interface for all (or selected) links of a network
in a single plot. A visual approach is better suited to
comprehend the large amount of data collected during
an experiment, providing for example an overview of
the connectivity in the network and an easy way to spot
connectivity problems. Coupled with traditional mea-
surements, such as a source node’s TCP sequence number graph, it can be easier to identify and understand the
spatial and temporal events that impact the performance.
We construct activity plots in the following way. During the experiment, all nodes record the successful reception of 802.11 frames, higher level events like route
changes, and application layer specific data. The background time stamping application permits to synchronize the local traces after the test run and to merge the
logs into a single experiment trace. From this trace file,
we re-extract node-specific reception events and classify
them according to their origin, creating a “link” view,
and distinguish different message type in a graphical
way. A link from node x to node y is denoted as x → y.
The links are laid out on the y-axis of the figure, while
the x-axis represents time. Different activity events on a
link are plotted as:
+ (large plus) unicast packet destined for node y.
(small plus) unicast packet overheard by node y.
(diamond) broadcast packet. AODV route requests
are highlighted with a large diamond.
We will now review some of our previous findings and
discuss them using activity plots.
to their destination (i.e., the TCP data flow is not interrupted).
The activity plot also shows that the UDP data flow
sent from node 2 is not received until time 11. It further
shows that the broadcast packets sent by node 2 (i.e.,
AODV HELLO messages), and received before time 0,
stops being received up until time 11. From the activity
plot we cannot conclude if node 2 was unsuccessful in
contention for the wireless channel, or if node 1 did not
receive the packets due to interference. If there would
have been another node within connectivity of node 2 at
that time, it might have overheard the packets. But from
this information, we can not be sure. Clear though, is
that the TCP flow affects the reception of HELLO messages. The route request visible at time 11 just before a
burst of traffic on link 2 → 1, indicates that node 2 experienced problems to setup a route to node 1. The burst
is caused by the transmission of buffered packets once a
route is discovered. After the UDP flow is established,
the packet interspacing of the overheard TCP data packets is increased. This is in line with the time-sequence
diagram shown in Figure 7.
Another interesting part of the Roaming node experiment is the route change from one hop to two hops
for the TCP flow around time 27. The activity plot in
Figure 9 shows all activity originating from node 3, as
perceived at node 0, 1, and 2 during the time when the
route change occurs. A change in activity on the different links indicates connectivity problems on one of the
Figure 8. Activity plot showing the UDP flow
starvation at the beginning of the experiment
as seen by node 1.
An activity plot from the beginning of the Roaming
node experiment from section 3.4 is shown in Figure 8.
The plot shows all link activity as seen by node 1. In
this scenario we know that node 3 is supposed to send
TCP data to node 0 which is acknowledging the data
packets starting at time 0. The activity plot shows that
node 1 indeed overhears this traffic from nodes 3 and
0. At time 3, there seems to be a short interruption of
the acknowledgments. A separate inspection of the link
3 → 1 however shows that these acknowledgments got
Figure 9. Activity plot showing all packets sent
by node 3 illustrating the route change from
one hop to two hops on the TCP link.
The route change going from a direct connection between node 3 and node 0, to a two hop route over the
intermediate node 1, between time 27 and 32, can also
be seen. The large plus symbols in Figure 9 indicate,
before time 27, unicast packets on the link 3 → 0 and
after time 32 on link 3 → 1.
Another interesting observation is that only broadcast
packets are received by node 0 on the link 3 → 0 after
time 27. After a while, also broadcast reception is impossible due to the increasing distance between nodes 3
and 0. The range where only broadcast but not unicast
packets can be received is referred to as gray zones [6].
The main effect of the gray zone is on AODV, which
during this time still assumes a valid route because it
can receive broadcast HELLO messages, although no
data traffic gets through. In the activity plot, it can be
seen that nodes 1 and 2 actually overhear TCP retransmissions from node 3 while node 0 is in the gray zone
of node 3. Only when broadcast reception stops will
AODV time out its route. Not until after that time will
TCP’s retransmission trigger AODV to send a route request, searching for a new route (visible, for example,
on the link 3 → 1 just before time 32). However, because of the exponential back off of TCP, this route discovery might not occur immediately after the gray zone.
The TCP retransmission can be in a long timeout, effectively delaying the route re-discovery. This stall can be
as long as twice the duration of traversing a gray zone,
and severely worsens the effect of the problem.
Discussion and Conclusion
This paper has studied the effects on a low rate multihop UDP flow from a competing TCP flow. It was done
through several experimental test scenarios where the
TCP flow interfered, shared hops with the UDP flow and
under mobility with dynamic routing. The purpose has
been to study the interactions between the protocols at
different layers.
The results indicate that although we use a moderate
rate UDP flow, TCP’s congestion control does not seem
efficient enough to only have marginal impact on the
other traffic in the network. When the two data flows do
not share common links, we observe increased packet
interspacing in the UDP flow, caused by jitter and to
some extent packet loss. Instabilities in the form of short
stalls are observed in TCP. Further analysis might show
if TCP retransmissions are caused by lost packets or the
high fluctuations in round trip time.
In the case where UDP and TCP share a common
link, contention is significantly higher resulting in increased UDP packet loss and more significant TCP interruptions.
Dynamic routing, in particular when using broadcast
neighbor sensing, adds another dimension of instability
to ad hoc networking. Hidden terminals and channel
capture effects cause otherwise stable routes to become
unstable, simply because routing control messages are
lost due to the competing data traffic. In our most complex scenario, the initial UDP flow experienced an average of 21.6% packet loss.
Our experiments show the complexity of the interactions in wireless multi-hop ad hoc networks. Therefore,
we have developed an analysis tool to graphically analyze interactions between the protocols at different layers and for different nodes.
The authors would like to thank Laura M. Feeney for
invaluable feedback on a prior version of this paper.
[1] The Uppsala University Ad Hoc Implementation Portal.
[2] G. Anastasi, E. Borgia, M. Conti, and E. Gregori. Wi-Fi
in Ad Hoc Mode: A Measurement Study. In Proceedings
PerCom, Orlando (Florida), 2004.
[3] M. Gerla, K. Tang, and R. Bagrodia. TCP performance
in wireless multi-hop networks. In Proceedings of IEEE
WMCSA’99 (to appear), (New Orleans, LA), February
[4] V. Gupta, S. V. Krishnamurthy, and M. Faloutsos. Improving the performance of TCP in the presence of interacting UDP flows in ad hoc networks. Unpublished.
[5] G. Holland and N. Vaidya. Analysis of TCP performance over mobile ad hoc networks. Wireless Networks,
(8):275–288, 2002.
[6] H. Lundgren, E. Nordström, and C. Tschudin. Coping
with communication gray zones in IEEE 802.11b based
ad hoc networks. In Proceedings of The Fifth ACM International Workshop On Wireless Mobile Multimedia
(WoWMoM), September 2002.
[7] E. Nordström, P. Gunningberg, and H. Lundgren. A
testbed and methodology for experimental evaluation of
wireless mobile ad hoc networks. In Proceedings of Tridentcom, February 2005.
[8] K. Xu, M. Gerla, L. Qi, and Y. Shu. Enhancing TCP
fairness in ad hoc wireless networks using neighborhood
red. In Proceedings of MobiCom’03, San Diego, USA,
September 2003.
[9] S. Xu and T. Saadawi. Does the ieee 802.11 mac protocol
work well in multihop wireless ad hoc networks? IEEE
Communications Magazine, 39(6), June 2001.
Hop of No Return:
Practical Limitations of Wireless Multi-Hop Networking
Marina Petrova, Lili Wu, Matthias Wellens, Petri Mähönen
Aachen University, Department of Wireless Networks, Kackertstrasse 9, 52072 Aachen, Germany
E-mail: [email protected]
In this paper we report in a concise form our results
from the measurement campaign that was performed
to test throughput efficiency of multi-hop 802.11 networks, including also heterogeneous networks with
802.11a, .11b, .11g, and Bluetooth 1.1. One of the
main contributions of the paper is that it is not reporting only simulations, but carefully monitored and calibrated measurements. The paper confirms the expectation that if relaying nodes have only one radio the
number of wireless hops should be strictly limited.
The wireless communication has emerged as a
mainstream paradigm during the last decade. The
number of wireless LANs, especially 802.11 based
systems a.k.a. Wi-Fi, has increased extremely rapidly,
and this development is, in part, driving the wireless
data communications development.
It is a common practice to use simulation tools
for analysing and verifying the behaviour of wireless networks as it is relatively fast way to obtain results. However, much remains to be done to make
simulations more robust and reliable. The simulation
based studies should also be validated with measurements at reasonable limits. In the case of WLANs,
most of the constraints in the reliability come from
the relatively simple PHY/MAC IEEE 802.11 implementation in many commonly used simulators [20].
Many earlier simulation studies have been focusing on
testing the interaction between TCP and the wireless
MAC protocols, evaluating the fairness and measur-
ing the throughput analysis of TCP in multi-hop IEEE
802.11b scenarios, both in indoor and outdoor environment. Besides the numerous theoretical and simulation based articles several practical studies have been
performed on wireless multi-hop networks [4, 10, 12,
18, 21].
The motivation of our work reported in this paper is to summarize the studies based on real testbed
measurements that we made recently on wireless ad
hoc networks. Several of them were done and reported about in other publications but doing a whole
measurement campaign in one environment allows detailed analysis and comparison of results accepting
that not all scenarios can be presented in full detail
here. Great care was taken to find out practical problems that could be encountered with real physical deployment of wireless multi-hop mesh networks. In
this paper we study the TCP and UDP performance
in indoor IEEE 802.11b and heterogeneous multi-hop
scenarios. The heterogeneous scenario combines devices with different wireless interfaces such as IEEE
802.11b/g and Bluetooth. Due to the limited range
of the wireless interfaces used, multiple hops are typically needed to enable the communications.
Our target scenario is mostly based on the idea
of mesh or community networks, i.e., most of the
wireless nodes are static or very slowly moving. We
present results from a comprehensive set of measurements and study issues such as TCP and UDP throughput and performance of the MANET routing protocols
(AODV and OLSR) in our scenario. Since the users
are mostly static (the topology of the network is not
rapidly changing), we make some comparative studies
using static and dynamic routing protocols. In addition, we also measure the difference in performance
between MANET routing protocols and OSPF (which
is widely used in fixed networks and could be a possibility also for wireless environment as long as the
topology is relatively static). This is motivated by the
recent activity towards a new version of OSPF including extensions for wireless networks [14].
The rest of the paper is structured as follows. Section 2 gives a brief overview of the relevant work on
wireless ad hoc networks outlining the constraints of
TCP over wireless, the performance of the mostly used
ad hoc routing protocols (AODV and OLSR) and the
fairness and capture effect issues. In section 3 we give
a short description of our testbed setups along with the
comparison and analysis of the results. We also discuss the constraints and limitations of the present ad
hoc networking solutions based on the measurement
studies. We conclude the paper in section 4 by outlining some potential future research avenues.
Related Work
In a large number of recent studies on WLANs and
ad hoc networks, authors have studied the performance
of TCP over IEEE 802.11. The ’misbehaviour’ of TCP
over wireless is a consequence of several issues, and is
well recognized problem [11, 13, 22].
First, as the topology of the ad hoc network can
change and some of the wireless links can break, TCP
will experience timeouts that lead to severe performance impact. TCP is a transport protocol that by
changing its window size adapts the transmission rate
to the available network bandwidth. As such, it functions successfully in the fixed networks. However, in
the wireless network a packet can be lost not only due
to congestion but also because of the errors in the wireless channel. Often link-layer contention in the case of
hidden terminal problem can cause additional transmission errors. No matter the type of the loss, it is
incorrectly interpreted as a sign of congestion, causing
adaptation of the TCP window size and reduction of
the data flow. Many efficient mechanisms have been
proposed for improving the TCP performance in wireless ad hoc networks, see, for example, [6, 22] and
references therein. Second, it is shown that TCP performance in an ad hoc multi-hop environment is sensitive to different parameters such as packet size and
TCP window size [11]. It can also be verified from
measurements that for a specific network topology and
traffic flow, there is a TCP window size at which the
throughput reaches the highest value. Further increase
of the window size does not lead to a better result. This
result is obvious, but it is not always properly taken
into account. The third issue causing performance
degradation is due to the interaction with the IEEE
802.11 MAC protocol. More specifically, when used
in the multi-hop networks the present IEEE 802.11
MAC protocol may cause unfairness between competing TCP traffic flows, and a capture of the whole wireless channel by a single node can occur relatively easy.
Although the system could benefit from different
protocol boosters or TCP modifications, we are leaving them strictly out from our study. We are limiting our measurement campaign to unmodified, offthe-shelf solutions, since these are building blocks that
are mostly used in testbeds, community networks, and
simulation studies.
The standard MANET routing protocols typically
find routes using the minimum hop-count metric.
Many studies have shown that this approach often selects paths with less capacity than the best possible
paths existing in the network [10]. The reason behind
that is the binary logic of the algorithm: in the process of choosing the paths it is assumed that the links
are either able to deliver packets or not. This simple
logic can be improved, if for example a link quality
information (SNIR) is used as a criteria for selecting a
high-quality link. An alternative solution is, e.g., expected transmission count metric (ETX), as suggested
in [9].
Another point to mention is that many MANET
routing protocols such as AODV, for example, do not
take into account the effects caused by the multi-rate
nature of the wireless standards. This implies an unreal assumption that all available links in the network
are equal. In IEEE 802.11b for example the broadcasting of the HELLO messages is always done at 1 Mbps
whereas the unicast data packets are sent at higher
rates (up to 11 Mbps). Naturally the transmissions at
lower bit rates, which in the case of WLANs correspond to different modulation and coding schemes, are
more robust and can reach further. While the HELLO
messages can be heard, the same does not apply for
the data packets. This is a reason for appearance of
the ’gray zones’ [15], locations where the data pack-
ets experience a great loss. Furthermore Anastasi et.
al. [4] discussed on the impact of different physical
layer modes on the communication range, confirming what one would expect from theoretical PHY-layer
The number of portable devices with different
build-in communication technologies is increasing
rapidly. There are already several wireless technologies that can be used to build heterogeneous wireless
networks. Naturally the necessity for interconnecting the devices using different technologies is becoming reality despite the obvious problems and performance constraints that such a network will face (e.g.
reliable routing, higher delay because of forwarding
between technologies, necessity for multiple interface
node for bridging between technologies etc.). It is surprising how little heterogeneous and multi-radio environments are studied. Also the issue of harmonizing link-interfaces requires more attention. However,
there are some activities on the domain of transparent
link-layer technologies, such as Universal Link Layer
API (ULLA) in GOLLUM-project [1], Generic Link
Layer approach (GLL) as proposed by Ericsson, and
earlier Wireless Adaptation Layer (WAL) suggestions.
RTS/CTS switched off.
TCP and UDP Measurements and Simulations in Multi-hop Environments
We shall begin with a 802.11b based multi-hop scenario where both TCP and UDP throughput are measured as a function of number of hops. Figure 1
shows the results of the multi-hop measurement campaign that was also used for setting up ns-2 simulations. Having in mind the shortcomings of ns-2, we
improved the 802.11 MAC module and included enhanced error modelling. The reader is referred to [20]
for further details. The measurements were taken using 802.11b WLAN PC-cards based on Intersil Prism2
chipset explicitly set to 11 Mbps PHY mode. Each
measurement is averaged over several trials. Static
routing was used in the measurements. On the other
hand, simulations were based on AODV, but as there
was no mobility the routing protocol did not have a
great effect on the simulation results as one would expect. The reader should note that we adapted the propagation model in the simulation to be similar to the
measurement environment.
Measurement Results
TCP Measurement
UDP Measurement
TCP Simulation
Throughput (kbps)
A common way to quickly estimate the performance of a specific wireless network is to measure the
throughput. By throughput we mean the actual transport layer payload without any headers successfully received per second. In this section we shall analyze the
TCP and UDP throughput in a homogeneous 802.11b
multi-hop setup, we compare the throughput performance of 802.11a, 802.11b and 802.11g, and discuss
the results from the channel allocation analysis and
fairness issues when 802.11b and 802.11g are working together. We also give an overview of the outcome
from the throughput measurements in a heterogeneous
802.11b, 802.11g and Bluetooth environment, and finally compare the performance of the AODV, OLSR
and OSPF routing protocols.
The measurement setup in the multi-hop case was
a simple string topology in office environment. All
measurements were performed with laptops running
Linux Kernel v2.4.26 and TCP NewReno with selective acknowledgement and enabled timestamps and
UDP Simulation
Number of hop
Figure 1. End-to-end TCP/UDP multi-hop
throughput measurements and simulations.
The large number of wireless hops, in a single-radio
per node case, is very inefficient, as throughput is lost
rapidly. This is unavoidable even in the perfect envi-
Throughput (Mbps)
ronment without transmission errors, delays, etc., as it
is inherent for single radio repeaters. It is even more
serious in the realistic Wi-Fi multi-hop environment. It
is clear from figure 1 that already three hops are quite
suboptimal for many purposes. Further increasing of
the number of hops will result in even lower throughput. In our tests there were no external nodes contending for the channel. However, at a public hot spot
other users will also produce interference, so the end
throughput would fluctuate more and experience further degradation.
Tschudin et. al. have also introduced the ad hoc
horizon [19] of about three hops for 802.11 networks
and TCP. Our measurements confirm their simulationbased results and extend their setup by taking into account UDP transmissions and higher data rates.
Distance (m)
Figure 2. TCP Throughput measured with
802.11a, 802.11b and 802.11g.
Comparing IEEE 802.11a, .11b and .11g
After measuring the performance of IEEE 802.11b
we also included 802.11a and 802.11g devices into
the testbed. Both technologies are based on Orthogonal Frequency Division Multiplexing (OFDM) modulation scheme in 5 GHz and 2.4 GHz frequency band
and support bit rates from 6 to 54 Mbps.
There has been some uncertainty on how 802.11a
compares to 802.11g. Both technologies are quite similar on PHY layer but use different frequencies. Naturally the radio implementations are at least slightly
different. In order to compare the performance of both
technologies in a real indoor environment, a set of
single-hop throughput measurements was performed.
For each technology, the end-to-end TCP throughput
was measured at 8 different points (range from 2 m
to 30 m) using a packet size of 1500 bytes. In addition to 802.11g (Prism GT chipsets) and 802.11a
(Atheros AR5211 chipsets) we made 802.11b measurements (Prism2 chipsets) for comparison.
Figure 2 shows the TCP throughput of 802.11a,
802.11b, and 802.11g as a function of distance. In
general this graph also shows the SNR-dependency
of the TCP-throughput since they are closely related.
Comparing the 802.11a and 802.11g curves, we can
identify three segments: At short distances with LoS
(line of sight), as expected, both technologies reach the
same maximum throughput of about 23 Mbps. When
the nodes are further away from each other (> 20 m),
with obstacles in between, both technologies switch to
the lowest bit rate to maintain the connection. In the
range in between, the throughput of 802.11g clearly
outperforms 802.11a. Due to the higher path loss of
802.11a, the physical layer mode was already switched
to more robust modulation and coding that led to lower
bit rate. On the other hand, the throughput of 802.11b
is rather stable. At some points (e.g. distance from
12 m to 20 m), 802.11b even outperforms 802.11a
cards. In the whole measurements the effect of the rate
adaptation is clearly visible.
It should be mentioned that the performance of the
devices depends on the deployed hardware. Wireless
LAN cards from different vendors, such as described
in [5], can have a distinct impact on the data throughput, although these differences will not change the
general trends.
In general, the measurements show that 802.11g
and 802.11b can benefit from smaller propagation path
loss, and consequently 802.11g has larger range than
802.11a at the same bit rate. The OFDM-modulation
is only beneficial for rather small distances where the
SNIR at the receiver is still high enough.
Impact of Channel Assignment on 802.11b
In case the density of WLAN hot spots is high, the
channel selection becomes an important issue. We
Contention between IEEE 802.11b and IEEE
IEEE 802.11g was designed to offer higher datarates still keeping the backward compatibility to the
most popular IEEE 802.11b. Therefore it also works
in the 2.4 GHz ISM-band and supports the basic CCKmodulation of 802.11b.
The interworking between well-known 802.11b devices and nodes using 802.11g PC-cards is an interesting research issue. Therefore we analyzed the contention between one 802.11b link and another 802.11g
link, both running simultaneously.
Figures 4 and 5 show both the TCP and UDP
Transmission A
Transmission B
Total Throughput
Throughput (Mbps)
address this problem by comparing a set of measurements done with different channel assignments for two
simultaneously running traffic flows. The hosts are
located close to each other. Two TCP transmissions
(the packet size is 1500 Bytes) run simultaneously for
20 s, and the average throughput is recorded. In each
measurement we changed the channel assignment, so
that the channel distance between the two transmissions corresponds to 0 (same channel) up to 5, having in mind that there are 11 channels available in the
802.11b standard. Note that the difference between
the channels is not in frequency domain but according to their numbering from 1 to 11. Figure 3 clearly
shows the dependency of the throughput on the channel selection. When both transmissions run in the same
channel, the total throughput is close to the maximum
throughput for a single transmission. This indicates
the CSMA scheme of 802.11 MAC can fairly resolve
the competition in the same channel without decreasing channel capacity. Since the frame collisions in the
neighbouring channels (1 or 2 channels apart) cannot
be effectively avoided by CSMA the total throughput
is lower than in the single channel case. When the
frequencies are assigned 4 channels apart the throughput is almost the same as in the non-overlapping channels case. The results show that neighbouring channels should be avoided when assigning frequencies although slightly overlapping channels could be used
in a dense environment. We have earlier proposed a
distributed channel assignment mechanism based on a
graph colouring algorithm that minimizes the interference and improves the performance [16].
Channel Distance (# of channels)
Figure 3. Comparison of different channel allocation schemes with 802.11b.
throughput measured at each single link as well as the
aggregate throughput of the two flows. As it can be
seen from figure 4, we started the second transmission
five seconds later. The result will not change if we
reverse the order of the transmissions. By comparing
the averaged throughput results during the contention
phase and the first or last five seconds one can calculate
that the 802.11g reaches ∼48% of its maximum TCP
throughput. In contrast 802.11b reaches only ∼26%
of its maximum. The same capture effect can be determined in the case of UDP traffic. As UDP does
not send acknowledgements the collision probability
of one transmission is lower so that the percentages
are higher (802.11g: ∼67% and 802.11b: ∼40%) but
the capture-effect can still be seen.
If both connections are not anymore configured to
the same channel but adjacent channels the general result will not change. As nodes working on different
frequencies do not sense each other’s transmissions
frame collisions become more probable and the total
throughput decreases. The unfair effect of 802.11g
capturing the channel is less severe but still present.
As 802.11g is based on OFDM modulation it is in
principle more robust against transmission errors. Furthermore the receivers are more sophisticated compared to the older 802.11b devices so that there is
a higher probability for the 802.11g interfaces to re-
Total Throughput
Throughput (Mbps)
Throughput (Mbps)
Total Throughput
Time (s)
Time (s)
Figure 4. Contention between 802.11b and
802.11g for TCP traffic.
Figure 5. Contention between 802.11b and
802.11g for UDP traffic.
ceive a frame although a frame collision occurs. The
802.11b may loose its transmission and therefore defers for an extended interframe space and doubles its
backoff window. The higher frame loss and the resulting higher average backoff period cause the smaller
802.11b throughput. As this capture effect depends
on the sensitivity of the device it may look different if
other WLAN cards are deployed.
rectly compares the heterogeneous results to the single BT link. Although we extended the connection by
adding an 802.11g as well as an 802.11b link the maximum throughput only decreases from 395 kbps to 360
kbps. As expected, the slow BT link forms a bottleneck for the whole transmission and the performance
of the end-to-end connection is only determined by
this bottleneck-link. We verified this behaviour using
a heterogeneous link consisting only of 802.11b and
802.11g that reached nearly the same performance as
a single 802.11b link. If both WLAN-links were using
non-overlapping channels the throughput would be as
fast as a single 802.11b link. However, if both links
were using the same channel the transmission would
not be much slower.
The heterogeneous multi-hop connections are more
or less limited to the performance of the slowest link
involved. This behaviour limits the usability and the
applications of such networks. However, having heterogeneous connections in the network can be used to,
e.g., extend the range of BT or interconnect technologies.
Heterogeneous technology scenario
For the heterogeneous scenario we extended our
measurement setup by using Bluetooth as another
wireless technology working in the 2.4 GHz ISMband. After measuring the single hop performance
we combined it with 802.11g and 802.11b to form a
heterogeneous 3-hop connection. We used static routing such as described for the homogeneous multi-hop
measurements and set the two WLAN links to different
channels to avoid interference between them. The interference between BT and WLAN was not avoidable
as we deployed devices based on BT v1.1 that do not
adapt their hopping sequences in order not to interfere
frequencies used by nearby WLANs.
For detailed analysis of the interference effects between BT and WLANs the reader is referred to [7].
We again generated TCP traffic using Iperf and
measured the end-to-end throughput. Figure 6 di-
Comparison of Routing Protocols
We used a quite simple three-hop string topology
as presented in subsection 3.1 to study routing protocol sensitivity. In order to introduce alternative paths
Throughput (kbps)
Throughput (kbps)
BT Link
BT-g-b Link
Time (s)
Packet_size (Byte)
Figure 7. Comparison of different routing protocols: AODV, OLSR and OSPF.
Figure 6. Comparison of a single BT link and a
heterogeneous 3-hop connection consisting
of BT, 802.11g and 802.11b.
we extended the topology by adding a second parallel
node at both intermediate hops. The final static scenario consists of 6 nodes using 802.11b interfaces.
Besides the well-known ad hoc routing protocols
AODV [17] and OLSR [8] representing the reactive
and proactive approach, we chose OSPF as a third
routing protocol. Since we were using a static topology we expected a relatively stable scenario and OSPF
is an interesting option because of its wide deployment
in the fixed networks enabling a much easier integration of existing wired infrastructure and new wireless
We measured the TCP end-to-end throughput with
packet sizes of 1500 bytes. Figure 7 shows the measured throughput over a period of 180 s.
Although we used a static topology all routing protocols reveal stability problems; OLSR even looses
the connection for several seconds. This leads to a
smaller average throughput compared to the measurements with static routing presented in subsection 3.1.
When comparing the three protocols in detail OSPF
performs worst. The unreliable nature of the wireless channels is the main reason as OSPF was not designed for such environments. However, the performance is not much worse and no connection breaks
occurred. The extended version of OSPF may be an
interesting option for future work. The used OLSR
implementation from UniK (Universitetsstudiene på
Kjeller) [2] reaches for several periods the maximum
possible throughput but suffers severe instability problems. The AODV implementation from Uppsala University [3] on the other hand does not reach the same
maximum throughput but is more stable overall. We
cannot draw a final conclusion of which protocol is
the best suited for static small-scale scenarios as none
reached sufficient stability, which is one of the main
desirable requirements for routing protocol in possible
commercial deployments.
The results from our experimental study show that
massive scale multi-hop networks based on unmodified 802.11 and TCP are not efficient. They can
be build, but there should be clear reason for paying the price of considerably lower performance. In
many consumer and civilian applications the justification might be difficult to find. However, in the limiting
small-scale use one could consider the possibility of
applying multi-hop techniques for extending the range
of an access point in infrastructure mode. Adding one
or two extra hops between the access point and the end
user device reduces the need for infrastructure and increases the flexibility of the network.
We would also like to stress that although the
WLANs are quite plug-and-play for single-hop consumer use, when one is building optimized multihop networks this is not the case. Especially when
one is programming lower level functionalities or tries
to read MAC/PHY-layer parameters the lack of well
documented interfaces makes it difficult. Besides,
MANET routing protocols do not reach sufficient stability even in small-scale static scenarios.
As WLANs become more widespread and heterogeneous (802.11b and 802.11g are today’s commercial standard)the shown capture effect of more modern transceiver hardware and the smart frequency allocation will become issues that require more serious
considerations when deploying these networks.
If one is considering strictly the present day technology without extensions, we recommend staying below four hops in order to keep system performance at
the reasonable level. As pointed out also by other authors, if one wants to build mobile and massive ad hoc
networks, unmodified 802.11 technology might be inadequate - at least in the case of single radio per host.
We would like to thank the European Union (6HOP,
MAGNET and GOLLUM-project) for providing partial funding for our work.
[1] The GOLLUM-project website, [Cited on: 27th April 2005], 2004.
[2] UniK OLSR Implementation,
[Cited on: 19th April 2005], 2004.
[3] Uppsala
on: 19th April 2005], 2004.
[4] G. Anastasi, E. Borgia, M. Conti, and E. Gregori. WiFi in Ad Hoc Mode: A Measurement Study. In Proc.
of PERCOM’04, March 2004.
[5] Atheros Communications. White Paper: 802.11 Wireless LAN Performance, USA, 2003.
[6] B. Bakshi, P. Krishna, N. Vaidya, and D. Pradhan.
Improving Performance of TCP over Wireless Networks. Technical Report 96-014, Texas A&M University, 1996.
[7] C. Chiasserini and R. Rao. Performance of IEEE
802.11 WLANs in a Bluetooth Environment. In Proc.
of WCNC, Sept. 2000.
[8] T. Clausen, P. Jacquet, A. Laouiti, P. Muhlethaler,
A. Qayyum, and L. Viennot. Optimized Link State
Routing Protocol. In Proc. of INMIC ’01, 2001.
[9] D. S. De Couto, D. Aguayo, J. Bicket, and R. Morris.
A high-throughput path metric for multi-hop wireless
routing. In Proc. of MobiCom ’03, 2003.
[10] D. S. J. De Couto, D. Aguayo, B. A. Chambers, and
R. Morris. Performance of Multihop Wireless Networks: Shortest Path is Not Enough. In Proc. of
HotNets-I, Oct. 2002.
[11] Z. Fu, P. Zerfos, H. Luo, S. Lu, L. Zhang, and
M. Gerla. The Impact of Multihop Wireless Channel
on TCP Throughput and Loss. In Proc. of INFOCOM
’03, March 2003.
[12] M. Gerla, R. Bagrodia, L. Zhang, K. Tang, and
L. Wang. TCP over Wireless Multihop Protocols:
Simulation and Experiments. In Proc. of ICC ’99.
[13] A. Gurtov and S. Floyd. Modeling wireless links
for transport protocols. ACM SIGCOMM Computer
Communication Review, 34(2):85–96, Apr. 2004.
[14] J. Ahrenholz, T. Henderson, P. Spagnolo, E. Baccelli,
T. Clausen, and P. Jacquet. OSPFv2 Wireless Interface Type, IETF draft, expires Nov. 2005 (work-inprogress), 2004.
[15] H. Lundgren, E. Nordström, and C. Tschudin. Coping with communication gray zones in IEEE 802.11b
based ad hoc networks. In Proc. of WoWMoM ’02,
[16] P. Mähönen, J. Riihijärvi, and M. Petrova. Automatic
Channel Allocation for Small Wireless Local Area
Networks using Graph Colouring Algorithm. In Proc.
of PIMRC 2004, Sept. 2004.
[17] C. Perkins and E. M. Royer. Ad-hoc On-Demand Distance Vector Routing. In Proc. of WMCSA ’99, 1999.
[18] L. Qiu, R. Chandra, K. Jain, and M. Mahdian. Optimizing the Placement of Integration Points in Multihop Wireless Networks. In Proc. of ICNP ’04, 2004.
[19] C. Tschudin and E. Osipov. Estimating the Ad Hoc
Horizon for TCP over IEEE 802.11 Networks. In
Proc. of Med-Hoc-Net ’04, June 2004.
[20] M. Wellens, M. Petrova, J. Riihijärvi, and
P. Mähönen. Building a Better Wireless Mousetrap:
Need for More Realism in Simulations. In Proc. of
WONS ’05, 2005.
[21] K. Xu, S. Bae, S. Lee, and M. Gerla. TCP Behaviour
across Multihop Wireless Networks and the Wired Internet. In Proc. of WoWMoM ’02, 2002.
[22] X. Xylomenos, G. C. Polyzos, P. Mähönen, and
M. Saaranen. TCP Performance Issues over Wireless
Links. IEEE Communication Magazine, Apr. 2001.
Deploying MANET Test-beds
Thoughts on Mobile Ad-hoc Network Testbeds
Wolfgang Kieß, Stephan Zalewski, Andreas Tarp, Martin Mauve
Heinrich-Heine University Düsseldorf
{stephan.zalewski, andreas.tarp}
Currently, several research groups seek to develop
appropriate tools and methodologies for real-world
experiments on mobile ad-hoc networks. We argue that
this should be done as a community effort rather than
as independent projects. Furthermore we present our
view on the functionality a good testbed should provide based on reports from other research groups as
well as on our own experiences. This paper aims at
stimulating a discussion, it is not meant to be a comprehensive specification of requirements or a detailed
The two most common instruments used to evaluate algorithms for mobile ad-hoc networks (MANETs)
are theoretical analysis and network simulation. Theoretical analysis provides fundamental insights into the
characteristics of the investigated approaches, simulation enables their exploration in a dynamic environment. Both methods require a significant level of abstraction to reduce the real-world complexity of mobility, radio propagation, and hardware. As it has been
discussed in [13], the direct transfer of findings from
simulations to real-world systems is thus not advisable. As a consequence, theoretical analysis and simulation should be complemented by real-world experiments.
A surprisingly large number of these real-world experiments have already been conducted, leading to results and insights that were not foreseen in simulation or theoretical analysis. Most of the experimenters
also had to face unforeseen difficulties. Nodes failed
during experiments which was only discovered after
the experiment, results were difficult to reproduce and
therefore work was unnecessarily duplicated by multiple work groups or the network showed unexplainable
behavior. We believe that these problems can be alleviated if future experiments satisfy three key requirements of scientific experimentation:
• Reproducibility. Independent research groups
must be able to reproduce the results of an experiment. For mobile ad-hoc networks, reproducibility is a significant challenge due to the complex
impact of radio propagation and node mobility on
the results of an experiment.
• Comprehension. A scientist conducting an experiment must be able to access all relevant information to comprehend and explain the results of
the experiment. There is a need for tools that collect information on different layers and combine
this information to allow a detailed analysis.
• Correctness. Any experiment may suffer from
broken tools, errors with the setup and problems
when conducting the experiment. While reproducibility and comprehension will most likely reveal these problems, it is vital to the efficiency of
a researcher to be able to verify whether any given
experiment has produced valid results. This can
be supported by an established methodology and
a selection of appropriate tools.
To address these issues, there are efforts underway that aim at the development of methodology and
testbeds to support systematic experiments in mobile
ad-hoc networks. One early example is the APE
project [14], introducing, among other things, virtual
metrics for node distance in order to increase reproducibility. Several other testbeds are currently being
worked on [19, 8]. All these efforts provide very valuable individual contributions. However, it is our belief
that a community effort is necessary to establish realworld experiments as the third standard instrument to
evaluate algorithms and protocols for MANETs. This
effort must include an analysis of the functionality expected from a “perfect” testbed as well as a discussion
of the methodology for conducting experiments. Ideally this will lead to the design of a common architecture where the individual research groups can contribute reusable building blocks for a testbed that is
supported by the ad-hoc networking community.
This paper is meant as our initial contribution to this
effort. It describes work in progress not final results. In
Section Two we present our view on a perfect testbed
for mobile ad-hoc networks. Section Three describes
some of our experiences with experimenting and Section Four concludes the paper.
The Perfect Testbed
In the following we assume that a testbed consists
of two key elements: a number of physical devices
(nodes) which may be moved around individually and
the software to support and conduct the experiments.
In general an experiment for MANETs can be divided in several phases: implementation, experiment
specification, node configuration, setup verification,
execution, and analysis. For each of these phases we
discuss how an ideal testbed could support them and
present the current state of the art in this area.
The first phase of an experiment is the implementation of the algorithm to be tested. A good testbed
will support this phase in three ways: it will 1) help
to minimize the work required for the implementation;
2) seek to reduce implementation errors; 3) encourage
interoperability between algorithms implemented and
evaluated by distinct research groups.
As a lot of algorithms will be initially analyzed by
means of simulation, reusing the simulation code instead of reimplementing it eases the workload and re-
duces the potential for errors. Thus tools allowing to
use the same code both for simulation and the real experiment are a vital component of a MANET testbed.
Additionally, the testbed should support the implementation by providing well tested libraries for common MANET functionalities such as beaconing or for
tasks such as tracing.
Encouraging interoperability is mainly a matter of
interfaces and methodology. A good testbed will specify concise interfaces and best-practice methods for integrating new functionality such that it can be reused
by other research groups. It will also support interoperability through a clean and simple architecture.
The concept of reusing simulation code for realworld experiments has already been utilized in some
projects, namely in SURAN [1], WINGS [5] and the
routing protocol evaluation presented in [23]. A particularly promising approach seems to be nsclick [16]
which allows to run a click router [12] inside the
ns-2 [17] network simulator.
There also exist libraries supporting the implementation of algorithms for ad-hoc networks. One class are
libraries such as the PICA API [2], designed to provide
platform-independency for the implementations using
them. Other libraries offer additional functionality like
neighbor discovery, flooding or packet buffering during route discovery. One example is the ad-hoc support
library [11].
Experiment Specification
After the implementation is complete, the experimenter specifies the scenario used for the evaluation.
In order to allow other research groups to verify the
results, the specification should be a complete description of the experiment made available as a file in a standardized format.
There exist at least two variants of scenarios, strict
and loose scenarios. In a strict scenario each node follows detailed instructions on when to perform which
action. Although a rigid description of the experiment
fosters repeatability, there are setups in which this is
not suitable, e.g., if experiments are run as background
tasks on devices primarily used for other purposes. In
this case, a scenario with loose descriptions of the services and actions able to adapt to the current state of
the node is better suited. In both cases a good testbed
will support the specification of an experiment through
a set of tools. At the very least these tools will allow
the planning of node mobility and the timing of events
via a graphical user interface.
The only widely known tool for the specification of
strict scenarios are the choreography descriptions for
the APE testbed [14]. To our knowledge there are no
tools that support the specification of loose scenarios.
Node Configuration
When the scenario is prepared, the nodes need to
be configured with the information required to run the
experiment. This includes the implementation of the
investigated algorithms as well as the specification of
the actions and the movements of each node. This step
mainly comprises the distribution of files and the configuration of nodes (e.g., setting of addresses), thus it
should be automated as much as possible.
The key to the autonomous configuration of the
nodes is the experiment specification. Since this specification contains any relevant information on how each
node should behave, a good testbed will be able to install the required software and perform the necessary
configuration based solely on this information. This
can either be done by directly distributing the specification to each node or it can require its “compilation” to gain configuration files that are specific for
each node.
If the nodes are physically accessible to the experimenter, the automatic file distribution can be provided
with simple means, e.g., through a one-hop download.
However, in loose scenarios the devices are normally
not available for a direct download and the required
files therefore have to be distributed to nodes that are
already in the field. One approach to do this is to let
the nodes distribute the required files amongst themselves, i.e., whenever two nodes come in radio range,
they will exchange information and files on the scheduled experiments.
Automatic configuration is a concept used in
APE [14], the PRNET project [9] provided a mechanism for the remote configuration of nodes.
Verification of the Setup
The actual execution of an experiment that uses a
strict scenario is extremely costly in terms of man-
power and time. A verification of the test setup and
the used hardware in the forefield of the experiment in
a controlled laboratory environment is therefore vital
and should be supported by the testbed.
The verification can be divided into tests involving
one or multiple devices. Single device tests allow to
avoid problems such as lack of memory, low battery
power or physical damages. Tests with multiple devices can reveal problems only occuring due to the interaction between devices. An important multiple device test which should always precede an experiment
is running the complete setup installed on real devices
under laboratory conditions. Although these artificial
conditions prevent the acquisition of quantitative results, the setup is not expensive, can be repeated easily,
and allows the isolation of errors.
There exist a number of approaches on how to create a multi-hop network topology for the in-lab verification, e.g signal attenuation [10] or MAC filters
that drop packets based on the MAC source address of
packets and the virtual position of the sender [15, 22].
A good testbed can use the position information in the
scenario file to compute the virtual distances and control the topology accordingly.
Support during the Experiment
The main phase of the experiment starts with the
distribution of the devices. Each experiment will most
likely consist of several runs in which the nodes move
around. Finally, the devices need to be collected and
the phase is concluded by collecting the tracefiles from
the devices. The main phase has some properties
which necessitate a dedicated support by the testbed:
1) The time in this phase is expensive, only an optimal usage of the experimental time makes experiments
economically feasible; 2) Repeatability of this phase
is crucial for a scientific evaluation; 3) All information available here is valuable; Therefore, the testbed
should support the experiment by optimizing the usage of experimental time, by guaranteeing, or at least
fostering, repeatability and by collecting detailed information on the nodes’ actions.
The usage of the experimental time can be optimized by automating tasks and by avoiding errors and
therefore unnecessary repetitions. As device distribution and collection are physical tasks, the potential
for automation is small, here. This is different with
tasks not requiring a direct (human) interaction like
the tracefile collection or the movement of the nodes
which can be automated by mounting the nodes on
robots [3]. A large optimization potential also lies
in the avoidance of the execution of erroneous experiments. By controlling that all nodes act within
the parameters specified in the scenario, the testbed
should assure that exactly the intended experiment is
The repeatability of an experiment is provided if it
is possible to rerun the same experiment such that the
relevant parameters in both runs have sufficiently similar values. There are two ways to support repeatable
experiments, either by comparing the parameters after an experiment to determine if it was a repetition of
a prior experiment or by steering the experiments to
ensure that these parameters lie within an acceptable
threshold. Open issues in this context are the determination of the relevant parameters and the question
if it is technically and economically possible to record
these parameters.
For all aspects mentioned so far it is crucial to trace
the data on the behavior of nodes and on external influences as completely as possible. This data can be
used for a detailed post-run analysis as well as for the
steering of the experiment. The data to be recorded involves packet-level traces, timing and positioning information, states of higher level protocols as well as
physical and MAC layer logging.
To steer the experiment, the experiment control
component of the testbed should continuously compare the actual values of the relevant parameters to
those specified in the scenario. The testbed therefore
should provide a method to specify and control boundaries for these parameters, soft boundaries like “position between x-5 and x+5” as well as hard boundaries
like “GPS daemon running”. In case some of these
boundaries are violated, the testbed can adjust the behavior of the node during the run or mark the run as
invalid. If the violation is severe, this can render the
whole experiment unusable and should therefore be
known right during the experiment. Thus, the testbed
should support the transmission of status information
to a central monitor station and it should also be possbile to remotely login to the affected node to correct
errors or alter the configuration.
The experiment monitoring and the remote login
necessitate the exchange of management messages between the nodes and the monitoring station, possibly
during the experiment. This counteracts a primary design goal of the testbed, i.e., minimizing the interference of the test equipment (both hard- and software)
with the experiment. A solution to this is the “out-ofband”-transmission of all management messages. This
can be achieved either via a separate network interfaces during the experiment, as payload of experiment
packets with a dummy payload or in the pauses between the single runs of an experiment using the tested
network itself. Although the last method does not allow to stop erroneous runs directly, this is not problematic if runs are short. Furthermore, it is more practical
than the other methods as the first method is more expensive and the second not always feasible.
The concept of monitoring the state of the network
during the experiment has been used in the PRNET
Network Monitoring [9], the SURAN Automated Network Manager [1], the CMU Position and Communication Tracking daemon [15] and ATMA [18]. For
some of the tracing tasks, there exist standard tools
like tcpdump [21] for packet level tracing or gpsd [6]
which can help to record the node position. A means
to synchronize the clocks of the nodes is the usage of
a combination of NTP and GPS. The concept of determining if an experiment is a repetition of another one
by hindsight has been used in the APE virtual mobility [14]. To our knowledge, there exists no experiment
control that adapts the nodes’ behavior dynamically to
changes in the environment to foster repeatability.
Postprocessing of the Experiment
The postprocessing can be divided into organizing the raw data gathered during the experiment and
analysing it. This phase should be governed by the
principle that the raw data is a valuable resource. It
needs to be documented, stored and published. Based
on this data, independent researchers must be able to
verify any conclusions that are drawn from the experiments. A good testbed will provide mechanisms to
ease the documentation, storage and publication of the
involved files. Furthermore it will provide or incorporate an extensible toolset for analysing the raw data.
The first postprocessing step is the structured, permanent storage of all raw data. As there are also a
lot of other files involved in the postprocessing such
as scenario files or tools, the testbed’s automatic file
handling should include these. One possibility is the
implementation of a file management framework that
defines interfaces to access, view, annotate, process,
and store these files. The organization of the raw data
is concluded by the documentation of the events and
conditions not recorded in the traces but perceived by
the human participants.
The tools used to process the raw data provide functionality for consistency checks, data analysis or for
enabling trace-based simulation. All these tools can be
used for multiple experiments, thus the testbed should
foster reusability to reduce work. If the tools are modular and reusable, this also increases reliability as results can be easily reproduced. Therefore the goal
has to be the creation of a reusable standard toolset
for the postprocessing of MANET experiments which
should be publically available, extensible and well
documented. The analysis tools should support the
different input formats of real-world traces and simulator traces as this allows a direct comparison of results
from both evaluation methods.
The file management as well as the toolset can be
combined with a server that is publically available. If
software tools and raw data are available in an open
repository, other researchers can access and reuse the
data and tools or verify the results. We are not aware
of a toolset or testbed that supports the properties described above.
Orthogonal Concepts
The previous sections show how a testbed can support the single steps of an experiment. Besides that,
there are general concepts such as error avoidance,
reduction of workload, portability, modularity and
benchmarking not limited to single steps.
Benchmarks provide references to improve comparability. One option to realize benchmarks for
MANETs are baseline protocols. A baseline protocol
must 1) have the best performance in its class, typically achieved through the use of “illegal” means such
as global knowledge; 2) be easy to implement; 3) be
easy to test; If the baseline protocol is included in the
experiment, it builds an upper bound for the performance of the whole class of algorithms in this special setting. Instead of giving absolute numbers for
the performance of an evaluated protocol, it can be expressed as percentage of the maximum. Candidates for
baseline routing protocols are MERIT [4] or the “bestcase” routing described in [22] which both use global
knowledge to make routing decisions.
Another form of benchmarks are standard tests and
standard measure values. Standard tests can be standard topologies such as chains or grids as well as tests
on the maximum load the network can support or the
load without data traffic. Standard measure values provide characteristics of a protocol. Examples are delay,
packet order, route length or loss rate.
To reduce the workload and the errors arising due to
reimplementing the same code multiple times, using
the same source code for simulation, emulation, and
real-world experiments is highly desirable. We call
this SER integration. An additional advantage of this
approach is the feedback that can be given between
simulation and experiment. SER integration should be
used for the evaluated algorithms as well as for scenario files and analysis tools. This means that the same
scenario file should be able to drive a simulation as
well as an experiment or that the output of both steps
can be processed with the same analysis script.
Portability of the testbed software is very important
since a test setup used for large MANET experiments
will consist of heterogenous devices. As the full number of devices is only needed for a short time, sharing
devices among workgroups and using devices like laptops or cell phones not especially bought for experimenting will be common in experiments. Furthermore
not all workgroups will buy the same devices, leading
to further heterogeneity. Apart from this, the product
live cycle of mobile devices is short. If a device is
bought today, it may be not possible to buy the same
product in a year. Thus, the testbed software must be
very portable to run on these different platforms.
Finally, it seems beneficial to employ a highly modular testbed architecture rather than a monolithic approach. This mainly is due to the large selection of
tools allready available, including standard software
such as tcpdump [21] or gpsd [6]. These tools should
be used rather than reimplemented. In addition modularity will allow the easy exchange of components to
enable a competition on the level of individual components rather than complete testbeds. As a consequence
we believe that the system should be only loosely coupled with some kind of “glue” combining these components to form the testbed. An example where this
concept has been used to some extend are the scripts
controlling APE [14].
Since 2002 we have conducted experiments on realworld vehicular ad-hoc networks within the context
of the FleetNet project [7]. The experiences gained
through these experiments and the problems we encountered motivated us to investigate how to improve
experimentation with real-world implementations of
ad-hoc networks in general.
In a first step we tried to repeat experiments conducted by other research groups to confirm their results. This turned out to be extraordinarily difficult
since there is almost always insufficient information
available to reproduce the described results. Key issues were the lack of information on the environment
(in particular regarding the setup of the experiment,
the radio characteristics and the connectivity between
nodes), no access to the raw data gathered during the
experiment and a vast heterogeneity in the tools to
setup, conduct and evaluate an experiment. These experiences led us to the first attributes of a good testbed:
all information, code and tools need to be published,
preferably in a standardized way.
We have conducted in- and outdoor experiments of
a simple flooding algorithm for static ad-hoc networks
with seven to ten nodes. The motivation for this very
simple setup was to isolate problems that would lead
us to general design criteria for experiments with mobile ad-hoc networks. In the following we sketch the
main observations, more details can be found in [20].
Inspired by [15], we started by measuring the radio
ranges of our hardware (iPAQ5550 IEEE 802.11b) and
discovered that the iPAQ radios were sometimes able
to successfully deliver ping packets over more than
900 m while already a tree in the line-of-sight between
two nodes can block a transmission. Thus, setting up a
reliable, reconstructible 7-node/6-hop string topology
for preparatory tests was only possible by carefully positioning each device around the edges of a building.
Our next step was to set up a multi-hop topology where every node had multiple neighbors. After
several measurement sessions, a suitable experimental
site seemed to be the university parking lot. Later on
we discovered some undesirable properties of this location. As the library and other university buildings
are close, there were other WLANs present requiring
the careful selection of the radio channel (a problem
also described in [18]). Another issue were moving
cars possibly leading to frequent changes in the topology even though the nodes themselves did not move.
To determine the connectivity between the nodes
at the start of an experiment, each node transmitted
a number of beacons. While one node transmits its
beacon at a time, all other nodes record the packet reception. The necessary exact coordination is difficult
in a distributed ad-hoc network but can be achieved
with a domino effect. Prior to the experiment a route is
determined that includes all nodes of the network. The
nodes use the sequence imposed by this route to coordinate their beaconing. The first node starts with its
beaconing. Its successor will take over once this node
has finished. This is repeated until the last node has
transmitted its beacons. The domino approach worked
well for our small networks, however it can be foreseen that a more sophisticated mechanism is required
for larger and more dynamic networks.
Based on the experiences gathered during the FleetNet project, we were aware of the problem of having to reimplement the algorithms when switching between simulations and experiments. Therefore, we implemented flooding in click [12]. With the help of
nsclick [16], the same code used for the experiment
can be run in ns-2 [17]. We used the integration for
two purposes: the first is to debug and gain first experiences with the implementation and the setup of the experiment in a controlled simulator environment. When
experiments are conducted without proper tests, this
can waste a huge amount of manpower: a bug in the
configuration of one of our in-building tests rendered
a whole series of test runs useless. A miscalculation of
the pause time between sending packets resulted in intermixed flooding attempts. A simulation prior to the
testing could have revealed this problem without incurring significant overhead. The second reason for an
integrated simulation and experimentation approach is
to use the data gathered during an experiment to im-
prove simulations with information about real-world
radio characteristics. To model the real setup as exactly as possible, also the positions of the nodes in the
real experiment must be known. We used GPS for this
purpose and encountered the well known problems of
position jitter and dependence on clear view to the sky.
During our experiments it often happened that either a link, the used software or a whole node failed.
Every time this happened, we had to check each node
manually. As a provisional solution, we implemented
a simple in-band one-hop status check. Each node
wrote its current status to a file accessible via HTTP.
To control the nodes’ status, it was sufficient to walk
around and use a perl script to retrieve the status file
from each node. Obviously, this approach has several limitations: it requires to walk in the radio range
of each node to be checked, the transmissions “contaminate” the experimental data, and the correction of
an error still requires physical access to the affected
node. Because of these experiences we believe that
node monitoring will be a very valuable element of a
good testbed.
Another issue appeared during the postprocessing
of the data from the experiments and the simulations.
As the output format of the simulator (ns-2) differs
from the trace format of the experiment (tcpdump),
each tool for the analysis of the results had to be implemented twice. Furthermore there exists currently no
good solution for commentation and documentation of
the raw data such that special events during the experiments can later on be remembered and reconstructed.
One key question occurred during the postprocessing: how good was the performance of the simple
flooding strategy? Of course there are absolute values on the number of transmitted packets and the rate
of received messages. However, this does not provide
any clue on how good our approach worked under the
given circumstances in comparison to an (illegal) optimal solution. Thus benchmarks would be a great help
in determining how well a given solution performed.
Conclusions and Outlook
We believe that a MANET testbed should be open
source, not restricted to special hardware, customizable, and not bound to any specific location. It must
support reproducible, comprehensive, and correct ex-
periments. While there are promising individual contributions towards this goal, currently no approach
fully satisfies these demands. For a multitude of reasons the best way to implement such a testbed seems
to be a joint effort of the MANET community. The
most important one may be a broad acceptance by the
A first step towards this goal should be a specification of the functionality that the testbed must provide.
We expect that this will be a controversial discussion
and hope that this paper may contribute to this effort.
Once the specification has become stable, the key factor to a successfull community effort will be the design of a highly modular system where each research
group can contribute individual parts. We expect that
the specification of scenarios, the scripting for running
experiments and the format of the raw experiment data
will require immediate attention and provide the glue
for the connection of the individual contributions. In
order to stimulate these first steps we have set up a wiki
[1] D. A. Beyer. Accomplishments of the DARPA
SURAN program. In Proceedings of MILCOM’90. IEEE, Sept 1990.
[2] C. T. Calafate, R. G. Garcia, and P. Manzoni. Optimizing the implementation of a MANET routing protocol in a heterogeneous environment. In
The Eighth IEEE Symposium on Computers and
Communications, June 2003.
[3] Emulab - mobile wireless networking.
[4] A. Farago and V. R. Syrotiuk. Merit: A unified framework for routing protocol assessment
in mobile ad hoc networks. In Proceedings of
MobiCom’01, pages 53–60. ACM Press, 2001.
[5] J. Garcia-Luna-Aceves. Wireless internet gateways (WINGs) for the internet. Technical report,
University of California, Santa Cruz, 2001.
[6] gpsd: a GPS service daemon.
[7] H. Hartenstein, B. Bochow, A. Ebner, M. Lott,
M. Radimirsch, and D. Vollmer. Position-aware
ad hoc wireless networks for inter-vehicle communications: The FleetNet project. In Proceedings of MobiHoc’01, Long Beach, California,
October 2001.
[8] S. Jadhav, T. Brown, S. Doshi, D. Henkel, and
R. Thekkekunnel. Lessons learned constructing
a wireless ad hoc network test bed. 1st Workshop
on Wireless Network Measurements (WINMee
2005), April 2005.
[9] J. Jubin and J. D. Turnow. The DARPA packet
radio network protocols. In Proceedings of the
IEEE, volume 75, pages 21–32, January 1987.
[10] G. Judd and P. Steenkiste. Repeatable and realistic wireless experimentation through physical
emulation. ACM SIGCOMM Computer Communication Review (CCR), 34(1):63–68, 2004.
[11] V. Kawadia, Y. Zhang, and B. Gupta. System
services for ad-hoc routing: Architecture, implementation and experiences. In Proceedings
of MobiSys’03, San Francisco, California, May
[12] E. Kohler, R. Morris, B. Chen, and J. Jannotti.
The Click Modular Router. ACM Transactions
on Computer Systems, 18(3):263–297, August
[13] D. Kotz, C. Newport, R. S. Gray, J. Liu, Y. Yuan,
and C. Elliott. Experimental evaluation of wireless simulation assumptions. In Proceedings of
MSWiM’04, pages 78–82, October 2004.
[14] H. Lundgren, D. Lundberg, J. Nielsen, E. Nordström, and C. Tschudin. A large-scale testbed
for reproducible Ad Hoc protocol evaluations. In
Proceedings of WCNC’02, pages 337–343, Orlando, FL, March 2002.
[15] D. A. Maltz, J. Broch, and D. B. Johnson. Experiences designing and building a multi-hop wireless ad hoc network testbed. Technical Report
CMU-CS-99-116, School of Computer Science,
Carnegie Mellon University, 1999.
[16] M. Neufeld, A. Jain, and D. Grunwald. Nsclick:
Bridging Network Simulation and Deployment.
In Proceedings of MSWiM’02, pages 74–81, Atlanta, Georgia, September 2002.
[17] The ns-2 network simulator.
[18] K. Ramachandran, K. Almeroth, and E. BeldingRoyer. A novel framework for the management
of large-scale wireless network testbeds. 1st
Workshop on Wireless Network Measurements
(WINMee 2005), April 2005.
[19] D. Raychaudhuri, I. Seskar, M. Ott, S. Ganu,
K. Ramachandran, H. Kremo, R. Siracusa,
H. Liu, and M. Singh. Overview of the ORBIT radio grid testbed for evaluation of nextgeneration wireless network protocols. In Proceedings of WCNC’05, New Orleans, LA, March
2005. IEEE. (to appear).
[20] A. Tarp. Experimental evaluation of flooding in
ad-hoc networks. Bachelor thesis, 2005. Department of Computer Science, University of Düsseldorf.
[21] tcpdump: a tool for network monitoring, protocol debugging and data acquisition.
[22] Y. Zhang and W. Li. An integrated environment for testing mobile ad-hoc networks. In Proceedings of MobiHoc’02, pages 104–111. ACM
Press, 2002.
[23] G. Zhou, T. He, S. Krishnamurthy, and J. A.
Stankovic. Impact of radio irregularity on wireless sensor networks. In Proceedings of MobiSys’04, pages 125–138. ACM Press, May
Experiences Deploying an Ad-hoc Network in an Urban Environment
Peter Barron, Stefan Weber, Siobhán Clarke, and Vinny Cahill
Distributed Systems Group,
Department of Computer Science,
Trinity College,
Dublin 2, Ireland.
email: {Peter.Barron, Stefan.Weber, Siobhan.Clarke, Vinny.Cahill}
Studies of mobile ad-hoc networks are in the most
part restricted to simulations and theory. They have,
to this point, rarely ventured into the real world on
a large enough scale to make significant statements
about their behavior or performance. The lack of evaluations is largely due to the practicable implications of
deploying and evaluating such networks in a real environment. In order to analyse the problems we have
built a Wireless Ad-hoc Network for Dublin (WAND).
The network provides a large-scale testbed for applications and protocols for mobile ad-hoc networks. It
offers the opportunity to explore the behavior and performance of a variety of routing protocols in a reallife environment and an ideal platform for investigating the use of mobile applications in an urban environment. In this paper we present WAND and the experiences gained in building such a network.
1. Introduction
Much of the research in ad-hoc networks has been
restricted to the evaluation of solutions through simulation. The simulators used typically aim to represent the different software and hardware components
within the system as well as the physical environment
in which they operate. While this provides a useful
method of validation the simplified assumptions made
of the physical environment limit the scope of what
can be achieve from them [4, 8]. For instance, the phenomena of gray-zones were only discovered through
experiments conducted real world environment. It is
therefore necessary to complemented simulation studies with real-world experiments to identify phenomena
that would otherwise go unnoticed within a simulator.
For these reasons, we have built WAND - the Wireless
Ad-hoc Network for Dublin - a large-scale testbed for
ad-hoc network protocols and applications.
The network covers the centre of Dublin along
a 2km route from Trinity College Dublin to Christ
Church Cathedral. The area is seeded with a number embedded devices with wireless connectivity.
These devices have been custom-built to meet the
needs of the WAND network. They are housed in
3x3x6inch containers that accommodate a stack of
PC/104 boards, IEEE 802.11b PCMCIA cards, and
two patch antennae. The embedded devices are hosted
on traffic lights and cameras along the route to provide a minimum level of connectivity. The embedded
devices form a sparse population of wireless network
nodes. This sparse coverage is constantly available
and the embedded devices can be configured to create a variety of network models. The testbed can be
further populated through the introduction of mobile
nodes such as laptops, PDAs, and other mobile devices
with wireless connectivity. The network provides researchers with a flexible platform for developing and
investigating different protocols and applications for
ad-hoc networks.
In designing the WAND network, the goal has been
to provide a platform that would allow basic research
on the behavior and performance of ad-hoc routing
protocols in a real-life environment. The choice of
an urban setting, while also providing a fertile en-
vironment for mobile applications, exposes the network to the problems and challenges of operating in
an environment that is inherently unpredictable. The
changes in weather, movement of people, and that of
cars, trucks, and buses, while the presence of buildings all have an effect. Rather than avoiding these
issues we aim to include and measure their effect on
the behaviour of the network and its components. The
WAND testbed also offers a unique opportunity to explore the different aspects of building applications for
wireless ad-hoc networks in an urban setting. The goal
has also been to use the WAND network to build integrated traffic systems, pervasive computing applications, and location-based services, and as such, it provides an excellent testbed for investigating these areas
of research.
In this paper we discuss the technology employed
in the test-bed and the experiences gained in deploying the WAND network. We then present some initial
results of the throughput over a series of nodes using
AODV [14].
Figure 2. Hardware used to WAND node.
The streets in the area are open to public traffic
which run on either one- or two-way streets. The vehicular traffic on these streets consists of a mixture
of cars and buses; articulated trucks do not enter this
area of the city centre. Most of the vehicles have
a height between 1.5m and 2.5m except for doubledecker buses which have a height of 4.40m.
2.1. Hardware
2. Description of WAND
The section provides details of the WAND testbed.
Describing the environment that it is operates in, the
hardware that is used to form the initial sparse population of fixed nodes for the network, and the software
that is installed on each of these nodes.
The testbed is locate at the heart of the shopping
and business districts of Dublin along a 2km route.
It is seeded with a number of custom-build wirelessenabled embedded PCs. The nodes are distributed
along the route in the form of a line at a distance of
about 200m apart. Each pair of neighbouring nodes is
installed within line of sight of each other. There are
currently eleven of these nodes located along the route.
The nodes are mounted on traffic lights and camera positions at a height of about 3m. Figure 1 provides an
overview of the area covered and the distribution of
the nodes. The GK node is used as gateway from the
WAND network to the college network.
The streets the network is deployed on are approximately 16-22 m wide and are lined by buildings of 4-6
stores high. The majority of these buildings have thick
stone-walls that block the propagation of radio signals
onto the neighbouring streets running in parallel.
The hardware for the nodes that form the initial
sparse population for the WAND network are custombuild embedded PCs. Each node consists of a stack
of PC104 boards [13, 12], a number of 2.4GHz PCMCIA wireless cards, an enclosure, and a set of antennae. The stack of PC104 boards comprises of a motherboard and two PCMCIA PCI-boards (see Figure 2).
These boards are mounted inside an enclosure together with a 30GB hard-drive. The motherboard employed in the system is a 400-700 MHz EEPD C3VE
board featuring a Pentium-class processor, 128MB of
memory, a video and an Ethernet adapter as well as an
IDE controller. Each of the PCMCIA boards have two
PCMCIA slots. These slots are filled by three CISCO
Aironet 350 cards which provided 802.11b connectivity. The cards have two sockets that connect to external antennae instead of an internal antennae which is
more commonly used. The sockets are connected via
10cm cables to RP-TNC connectors that are mounted
on top of the enclosure. The setup allows a variety
of antennae to be used by the nodes. At the time of
writing, each of the fixed nodes is equipped with two
patch-panel antennae with a gain of 8.5 dBi. The radiation pattern of these antennae is in the shape of a
Figure 1. Map of the area covered by WAND.
hemisphere in front of the antennae. The patch-panel
antennae of neighbouring nodes are directed towards
each other.
2.2. Software
The platform provided by the nodes is similar to
what your might find on a Pentium based PC with
a number of extra wireless network cards connected.
At present, the nodes have been installed with RedHat Linux 9. The installation includes packages such
as web-server, a JDK installation (version 1.4.2), a
secure-shell daemon, plus the majority of the other
packages that typically come with this distribution.
The kernel, that by default comes with the RedHat Linux 9 installation, has been replaced with the
kernel from the RedHat Linux 7.3 distribution. This
is due to the RedHat 9, kernel version 2.4.20, using
PCI-hotplug which conflicts with the current configuration of the PC104 stack. The kernel that comes with
the distribution of RedHat Linux 7.3, kernel version
2.4.18-3 was developed prior to the introduction of
PCI-hotplug support in the kernel and operates without problems on the PC104 stack.
Currently, the OLSR [7], and AODV [14] routing
protocols have been installed on the nodes. The AODV
implementation - Kernel AODV (version 2.1) - is from
the National Institute of Standards and Technology and
the OLSR implementation is from [3]. It is possible to
install other protocols though at this stage we have not
done so.
2.3. Configuration scenarios
The aim has been to provide a flexible platform that
can support various network topologies and installation of different routing protocols. It is possible to
configure the sparse network of nodes, that make up
the WAND network, to provide a backbone and serve
as regular access points to mobile nodes. Alternatively,
they can also be configured to use the wireless cards
and antennae they have to resemble a network of adhoc nodes. Mobile nodes can then be connected to
form larger ad hoc networks.
The fixed nodes also provide a excellent platform
for running different pervasive computing, mobile applications, or integrated traffic systems scenarios. The
embedded devices used are ideally placed in the middle of the shopping, tourist, and business districts of
Dublin and have enough power to support applications that require quality-of-service support or extra
processing power. The fixed installation of nodes also
ensures that a permanent presence can be maintained
for example for location-aware services.
3. Description of Experiences
This section describes some the experiences gained
during the installation and preliminary evaluation and
use of the testbed. The issues listed are broken into
two categories. Those that occurred during the initial
installation and those that subsequently arose from the
day-to-day use of the network as a testbed for ad-hoc
routing protocols, and for application development and
3.1. Installation issues
The installation of the testbed involved an initial design phase followed by the deployment of the nodes
between Trinity College Dublin and Christ Church
Cathedral. The choice of hardware and software used
was based on an initial set of requirements for providing a flexible testbed for basic research into protocols and applications is a wireless ad-hoc network.
The deployment of the nodes was achieved in collaboration with the local authority who provided access to
the traffic lights and camera positions along the length
of the route. The experiences gained from building
WAND have shown to us many of the difficulties facing those looking to develop these types of networks.
In this we wish to highlight a number of them.
Hardware is difficult. During the initial design of
the testbed a number of hardware configurations were
considered. One of the major factors that influenced
the design was the availability of off-the-shelf hardware for IEEE 802.11a/b/g network cards. Most of the
wireless network cards either available in either PCMCIA or cardbus format. Thus the hardware of a fixed
node had to support the PCMCIA format.
A number of community wireless projects use PCBs
with one PCMCIA slot. One of the intend use for the
testbed is the development and evaluation of multimedia applications and quality-of-service (QoS) protocols. The specifications for these uses required more
processing power and storage capacity than the available PCB could provide.
We finally choose the current hardware configuration because of its compatibility to desktop PCs,
availability of off-the-shelf components and the ability to configure it to suit the required number of network cards. However, the lack of previous experience with this hardware platform and the lack of maturity of the hardware at the time lead to a number of
problems: Motherboards that comply with the PC104
standard and are based on Intel 80386 processors are
relatively mature and stable. High-end motherboards
that comply with the PC104+ standard and feature
Pentium-level processors are more complex. The PCIbus that was introduced with the PC104+standard requires components in a PC104+ stack to accurately
synchronize the timing of signals on the bus. This re-
quirement makes the use of PC104+ stacks more complex and difficult compared to PC104 stacks.
Positioning of nodes. When deploying the network
between Trinity College Dublin and Christ Church
Cathedral it was necessary to find the best positions
along the route for the nodes. However, we found that
there where a number of issues that restricted where
we positioned the nodes. First, was the actual range
that could be obtained from the nodes and the patch antennae being used. It was found that the distance could
vary considerably depending on the buildings, density
of vegetation, road congestion, and interference from
existing networks in the area. However, the biggest influence was the actual route taken by the network in
combination with height of the buildings in the surrounding area. In many cases the line of sight between
nodes was lost as the network weaved it way through
the city centre. The combination of factors significantly decrease what could normally be achieved when
line of sight could be maintained. Sources of power
for the nodes also provided to be an issue that restricted where we could place the nodes. The only
power sources available to us were the traffic lights and
some of the camera positions locate along the route.
This provided us with 17 possible locations for 11
WAND nodes along a 2km route. However, we found
a number of issues that limited the number of possible positions/locations for the nodes. The use of other
sources such as telephone kiosks, apartments, shops
would have enabled a better distribution of the nodes
and provided better connectivity; unfortunately these
locations were not available to us at the time of the
Interference from other sources. The location chosen for the network, while providing an excellent
testbed for mobile and pervasive computing research,
is in the middle Dublin City Centre at the heart of the
business and shopping districts. A wireless survey, using kismet, showed a large number of networks, approximately 35, already existing in the area. The majority were infrastructural based but there were also
a couple ad-hoc networks. To annoyed interference
with these networks it was necessary to choose a channel not being use by these networks. Unfortunately,
due the size of the WAND network the survey showed
that the majority of channels were already in use. It
was therefore necessary to pick a frequency that would
cause the lease amount of interference with the existing networks to ensure the WAND network could operate successfully.
Because of the unlicensed nature of the 2.4GHz
band this will increasingly become a problem for deploying these types of networks in urban environments. There is also number of other sources of interference that can cause problems and which are not
so easily detected. An example for one such source
is the coordination of traffic cameras that are installed
at various points along the route. These cameras are
controlled generally through a wired connection but
sometimes the last 100m to a camera pole is based on
wireless communication in the 2.4GHz.
Height of installation The height of the installation
of the fixed nodes is a compromise between security of
the hardware installation, the quality of the RF signal
and the accessibility of the hardware for maintenance.
One of the initial concerns regarding the height of the
installation was the clearance of the Fresnel zone. A
rule of thumb determines that 60% of the 1st Fresnel
zone should be free of obstacles. This meant that the
nodes needed to be installed at least 1.5m above every
obstacle given that the fixed nodes were to be installed
every 200m along the route. An installation height of
4m would have been ideal to locate the direct line of
sight in sufficient space over the heads of pedestrians
and to protect the nodes against possible vandalism.
However, this height was deemed to be inaccessible
for maintenance purposes. It was predicted that due
to the exploratory nature of the installation the nodes
needed to be accessible without great difficulty and so
a compromise was made to install the nodes at a height
of 3m. This height allows access to the nodes by ladders and gives enough clearance of the Fresnel zone
to provide a good signal. One of the main obstacles
due to the height of the installation is the interference
of buses. Double-decker buses have a height of 4.20m
and tend to block the line of sight between two fixed
nodes completely.
3.2. Maintenance issues
A number of projects and experiments were run following the deployment of the testbed. The experiences
gained from running them on the WAND network are
described within the following sections.
Configuration of services The wide distribution of
the WAND nodes along with the dynamic state of the
network has lead to difficulties in maintaining the configuration of the network in a sustainable way. Currently, access to the nodes is achieved either from a
gateway node located at one end of the network, or
via other wireless devices - laptops, PDAs - within the
WAND network itself. Maintenance and the continuing configuration of the WAND nodes can only be accomplished through either of these routes. However,
the preferred route of administration has to be through
the gateway node as updating nodes from mobile devices within the network itself is impracticable over a
sustained period. The problem we have found is that
all the nodes are not always accessible from the gateway node due to a variety issues. Therefore, configuring or updating services or the network configuration
is difficult to achieve throughout the network without
it failing. So far, we have not been able to find a satisfactory mechanism that would allow updates to be
propagated through the network and be applied unilaterally.
Development of applications A number of projects
have been based on the testbed. Cocoa [2], Visualisation, Multimedia between mobile devices using
OLSR. The use of the WAND network has proven to
be a valuable resource in understanding how to build
these types of applications. The sparse network of
nodes provides a stable level of connectivity for these
projects that could not otherwise be achieved without
a greater population of mobile nodes within the environment. However, the deployment of these experiments and the retrieval of results from the nodes has
proven to be challenging. Again, the accessibility to
all the nodes from one location has proven to be the
main problem. To provide a usable testbed it is necessary to be able to coordinate and configure the network
from one set of experiments to another. To be able to
Figure 3. Throughput over two hops.
deploy, configure, and retrieve the results in an effective manner. For the WAND network this has proven
to difficult to achieve due to the unpredictable nature
of the network itself.
Figure 4. Throughput between individual
pairs of nodes.
4.1. Throughput
This setup with no interference through traffic or obstacles in the line of sight results in high throughput
with little variance. The next set of nodes is located on
the same road, slightly further apart. The throughput
between these two nodes is equally good to the previous pair despite the increased distance; however, a
junction before node 3 where a large number of buses
turn off from the main road leads to large variation in
the throughput. The line of sight of last pair of nodes is
again disrupted by a bend in the road and the throughput shows a similar pattern to the throughput of the
second pair.
The graph in figure 4 shows the throughput between
individual pairs of nodes. In order to understand this
output correctly it is necessary to consider at the location of the individual nodes. The first pair of nodes
has direct line of sight which results in good throughput of around 3.5 MB/s. However, the nodes are installed on different sides of the streets and the traffic
on the street causes interference and leads to a variation in the throughput. The line of sight between the
next pair of nodes is interrupted by a bend in the road.
This causes a slight drop in the throughput between the
two nodes; additionally the traffic on this street causes
the throughput to fluctuate as well. The third pair of
nodes is located on the same side of a straight street.
The second graph (figure 3) shows the throughput
over a 2-hop chain. The first chain involves a pair of
nodes with a clear line of sight and a pair of nodes
that a separated by a bend in the road. This results
in relatively low throughput with great variance. The
next chain has the pair of nodes separated by a bend
as first pair followed by a pair that has a clear line of
sight. This results in less throughput than in the first
case with similar variance. The third chain has two
pairs of nodes with a clear line of sight; however one
of the pairs is separated by the junction with a high
amount of bus traffic that was mentioned earlier. This
setup results in high throughput but with great variance
due to the bus traffic. The fourth chain is similar to the
first chain in that the first pair in this chain has a clear
4. Evaluation
This section describes a number of preliminary
experiments used to examine the throughput of the
WAND network using the AODV implementation
from the National Institute of Standards and Technology over a multi-hop network.
line of sight - sometimes interrupted by traffic and the
second pair is separated by a bend in the road. This
again results in low throughput with great variance.
From these two experiments we can see first indications of the influence of the traffic on the roads on the
throughput. Also, these measurements show the effect
of the obstructions in the line of sight and its influence
on the throughput in 2-hop chains.
The influence of the number of environmental factors such as weather, time of day, etc has not yet been
investigated and requires further measurements.
4.2. Related Work
The work related to the experience reported here
can be categorized into 3 general areas: reports on
community networks, reports on testbeds targeting research in mobile ad hoc network (MANET) protocols
and reports on testbeds targeting the investigation of
sensor networks.
Community networks such as MIT’s Roofnet [1]
employ similar hardware to the one used for WAND
nodes. Roofnet provides an experimental 802.11b/g
mesh network which offers broadband Internet access
to participants. The nodes in this network are fixed
in a place and consist of small PCs with a wireless
and a wired interface. The wireless interface is generally connected to a roof-mounted omni-directional
antenna. The nodes provide routing between mobile
nodes and the internet. Roofnet provides a platform
for experiments to understand the nature of large-scale
wireless networks. The research focus of this network
is the study of link-level characteristics of 802.11 and
the development of high-throughput routes.
A number of research projects have proposed setups
of testbeds in order to evaluate protocols for MANETs
[5]. Maltz et al [10, 11] report their experiences during
an experimental evaluation of Dynamic Source Routing (DSR) protocol under the influence of a high-rate
of topology changes. The testbed for this evaluation
consisted of 8 nodes that comprised IBM Thinkpads
560X notebooks, a 900 MHz WaveLAN-I radio with a
6db omni-directional antenna and a Trimble 7400 GPS
receiver. The nodes were mounted in cars that were
driven in a set course at an urban road. By evaluating
the routing protocol in a testbed the researchers developed a number of new ideas for the improvement of
the protocol.
Tschudin et al [16] developed the Ad hoc Protocol
Evaluation (APE) testbed. This testbed is based on
a software package in form of a Linux distribution.
Laptops can be booted from a CD using this package. Tschudin et al report on experiments with up to
37 nodes and 3 routing protocols [9]. The evaluation
of their implementation of the Ad hoc On-Demand
Distance Vector (AODV) routing protocol in a realworld environment led to the discovery of gray-zones.
In these gray-zones a routing protocol would report a
valid route to a destination but almost no data would be
delivered to the destination. This phenomenon was not
encountered in simulations of the same routing protocol and provides one of the many motivations for evaluations of protocols in real-world scenarios.
The third area that is related to the experience reported here is the area of sensor networks. Ganesan et
al [6] describe the experience they gained while evaluating a sensor network of 185 nodes in an open park
area. The nodes are based on a 4MHz ATMEL ATmega with a 900MHz low power radio. They found
that while current RF-propagation models can be used
to model certain behaviour such as that of sparse wireless networks; they need to be extended for dense deployments such as dense sensor networks. Ritter et al
[15] describe a similar network based on a combination of a microcontroller and a number of wireless interfaces such as 433MHz RF and Bluetooth modules.
This envisioned application domains for this network
are ad hoc gaming and home automation.
5. Conclusion
In this paper we have presented a testbed for the development and evaluation of ad hoc network protocols
and applications within an urban environment. The experiences that we have gathered through the development and deployment of this testbed are various. It
is extremely difficult to design a complex and flexible
testbed that delivers reproduceable yet realistic results.
However, we believe that no simulation is able to
reflect the inherent uncertainties and erratic nature of
real-world environments. The behavior predicted by
simulation may vary dramatically from that observed
in real networks such as WAND. Thus the measurements done in such a testbed will complement our un-
derstanding of the behavior and performance of protocols in realistic environments and the appropriate design and implementation of applications for such areas
as pervasive computing.
6. Acknowledgments
The work described in this paper has been partly
supported by the Irish Higher Education Authority under the Carmen collaboration between Media Lab Europe and Trinity College. The authors are also grateful
to Dublin City Council for supporting the installation
of the WAND nodes and for Jade Mastersons for the
use of the map of Dublin.
[1] D. Aguayo, J. Bicket, S. Biswas, G. Judd, and R. Morris. Link-level measurements from an 802.11b mesh
network. In Proceedings of the ACM SIGCOMM’04
Conference, Portland, Oregon, August 2004.
[2] P. Barron and V. Cahill. Using stigmergy to coordinate pervasive computing environments. In Sixth
IEEE Workshop on Mobile Computing Systems and
Applications (WMCSA’04), pages 62–71, December
[3] C. Calafate, R. Garcia, and P. Manzoni. Optimizing the implementation of a manet routing protocol
in a heterogeneous environment. In Proceedings of
the Eighth IEEE Symposium on Computers and Communications (ISCC’2003), Kemer - Antalya, Turkey,
[4] D. Cavin, Y. Sasson, and A. Schiper. On the accuracy of manet simulators. In Proceedings of the Workshop on Principles of Mobile Computing (POMC’02),
pages 38–43. ACM, Oct. 2002.
[5] P. De, A. Raniwala, S. Sharma, and T. cker Chiueh.
Mint: A miniatuirized network testbed for mobile
wireless research. In Proceedings of IEEE Infocom
2005, Miami, Florida, USA, March 2005. IEEE.
[6] D. Ganesan, B. Krishnamachari, A. Woo, D. Culler,
D. Estrin, and S. Wicker. Complex behavior at scale:
An experimental study of low-power wireless sensor networks. Technical Report UCLA/CSD-TR 020013, University of California Los Angeles, 2002.
[7] P. Jacquet, P. Muehlethaler, T. Clausen, A. Laouiti,
A. Qayyum, and L. Viennot. Optimized link state
routing protocol. In IEEE International Multi Topic
Conference INMIC 2001, Lahore, Pakistan, 2001.
[8] D. Kotz, C. Newport, R. S. Gray, J. L. andYougu
Yuan, and C. Elliott. Experimental evaluation of
wireless simulation assumptions. In Proceedings of
the ACM/IEEE International Symposium on Modeling,Analysis and Simulation of Wireless and Mobile
Systems (MSWiM), pages 78–82, October 2004.
H. Lundgren, D. Lundberg, J. Nielsen, E. Nordstroem, and C. Tschudin. A large-scale testbed for
reproducible ad hoc protocol evaluations. In Proceedings of 3rd annual IEEE Wireless Communications
and Networking Conference (WCNC 2002), pages
412–418, Orlando, Florida, USA, March 2002. IEEE.
D. Maltz, J. Broch, and D. Johnson. Experiences
designing and building a multi-hop wireless ad hoc
network testbed. Technical Report CMU-CS-99-116,
Carnegie Mellon University, March 1999.
D. Maltz, J. Broch, and D. Johnson. Lessons from a
full-scale multi-hop wireless ad hoc network testbed.
IEEE Personal Communications, 8(1):8–15, February
C. PC/104 Embedded Consortium, San Jose. Pc/104plus specification version 1.2, August 2001.
C. PC/104 Embedded Consortium, San Jose. Pc/104
specification version 2.5, November 2003.
C. E. Perkins and E. M. Royer. Ad hoc on-demand
distance vector routing. In 2nd IEEE Workshop on
Mobile Computing Systems and Applications, pages
90–100, February 1999.
H. Ritter, T. Voigt, M. Tian, and J. Schiller. A
highly flexible testbed for studies of ad-hoc network behaviour. In Proceedings of International
Workshop on Wireless Local Networks (WLN2003),
Bonn/Koenigswinter, Germany, October 2003.
C. Tschudin, P. Gunningberg, H. Lundgren, and
E. Nordstroem. Lessons from experimental manet
research. Elsevier Ad Hoc Networks Journal, pages
221–233, March 2005.
MeshDV: A Distance Vector mobility-tolerant
routing protocol for Wireless Mesh Networks
Luigi Iannone
Serge Fdida
LIP6/CNRS – Université Paris VI – Paris – France
Abstract— In this paper we propose an overview of
MeshDV, a routing protocol for wireless mesh networks.
MeshDV combines proactive route computation for routers
and on-demand path setup for clients. This design choice
eases the management of client mobility and reduces
routing table size. Proactive route computation is performed using a Distance Vector approach. On-demand
path setup is obtained through new mechanisms that take
advantage of particular characteristics of mesh networks.
Here we focus on these new mechanisms, giving details
about algorithms, packet format, and packet exchange
procedures. We also present the main components of the
MeshDV prototype currently under development at the LIP6
laboratory of the University of Paris VI.
Mesh subnetwork
Client subnetwork
Fig. 1.
An example of Wireless Mesh Network deployment.
The natural evolution of wireless networks enables
new ways for mobile users to get connected to each
other. In this context, Wireless Mesh Networks (WMNs)
are a promising emerging technology. In WMNs, nodes
are composed of wireless mesh routers (WMR) and
mesh clients.1 The architecture of WMNs is two-tier,
composed of a subnetwork for clients and a subnetwork
for mesh routers. The client subnetwork has the purpose
of offering wireless access to any client allowed to
access the network. The mesh routers subnetwork is
the backbone used to route packets between clients
associated with different WMRs or between clients and
gateways. The gateways between the client and the mesh
subnetworks are the wireless mesh routers. Figure 1
shows an example of this type of mesh network.
Routing in such a complex and dynamic structure
exhibits different, still unsolved, technical challenges.
Despite the large amount of research performed, few
works have been actually implemented. Typical proposals simply apply to WMNs existing solutions, proposed
in other contexts. For instance, LocustWorld [3] uses
a standard linux-based AODV implementation, and the
Roofnet Project at MIT [4] uses a network address
For simplicity of presentation we will call mesh clients merely
translator (NAT) on each WMR. The management of
client mobility can be done either through mobile IP or
by embedding routing features in the clients. Mobile IP
has drawbacks, like triangular routing, which makes it
not suitable for WMNs. The second solution works with
a flat, ad hoc architecture and limits the access solely to
clients able to support routing.
MeshDV, our proposed architecture, is designed to
take full advantage of WMN architecture. MeshDV mixes
proactive route computation for routers and on-demand
path setup for clients. This design eases the management of client mobility and reduces routing table size.
Proactive route computation is performed using a Distance Vector (DV) approach. On-demand path setup is
performed by new extensions introduced in the routing
protocol in order to take advantage of the architecture of
In MeshDV, the backbone becomes totally transparent
to the clients, which do not need to embed any new
feature. This means that if two clients associated to
different WMRs wish to communicate, the set of WMRs
will forward the traffic at the IP level and not at the
link level. To obtain such a behavior, WMRs have to be
more than a simple forwarder and more than a simple
router. In particular, on the client subnetwork interface,
User Space
IPv6 Forwarder
Routing Socket
Client Manager
UDP6 Socket
IPv6 Raw Socket
NDP Proxy
ICMPv6 Socket
Pcap − Filter
ether dst wlanIetherAddr
and ip6 and not icmp
PF − Rules
Fig. 2.
pass out all on wlanI
block in all on wlanI
pass in quick on wlanI inet6 proto icmp6
pass quick on wlanI inet6 to localhost
Kernel Space
WMR functional architecture.
the WMR acts in such a way to let local clients think
that remote clients are in the local WLAN. Then, it is
up to the WMRs to find out to which WMR the client
is associated and route the packets accordingly.
The University of Paris VI is planning to cover the
whole university campus with a large IPv6-based wireless mesh network. MeshDV represents the prototyping
and research work developed at LIP6, for the deployment
of a mesh routing protocol.
The remainder of the paper is organized as follows.
In section II, the general architecture of MeshDV and the
communication setup are presented. In sections III to VI,
the main modules of MeshDV are described. Details about
the deployed testbed, and some preliminary results are
given in section VII.
The fundamental blocks of WMNs are WMRs, whose
functional architecture is depicted in Figure 2. Each
WMR has at least two wireless interfaces, one for the
client subnetwork and one for the mesh subnetwork.
We call the first interface wlanI, since it implements a
WLAN offering access to clients, and the second one
wmrI, since it permits communications between WMRs.
Next we describe the architecture of MeshDV and then
we show how communication are set up.
A. MeshDV Architecture
The general architecture of MeshDV is also presented
in Figure 2; dotted lines represent the data flows between
the two interfaces inside the WMR. Each module of the
presented architecture has a specific meaning and task.
The basic task of the Enhanced DV module is to
maintain proactively a route toward any other WMR
present in the backbone subnetwork. This Enhanced DV
module is an IPv6 [5] implementation of DSDV [1],
based on existing IPv4 code [2]. Some enhancements
have been introduced in order to cooperate with the
Client Manager module. This module is the only one
that accesses the kernel routing table.
The Client Manager module is responsible for discovering which client is associated to which WMR in
an on-demand fashion. When a local client wishes to
communicate to a remote client (associated to another
WMR), it sends a multicast request in order to discover
where the client is.2 The result of this query is stored
in the Foreign Client Table (FCTable). The Client
Manager module also manages the reply to requests sent
by other WMRs. Furthermore, since MeshDV must be
aware of the clients that are associated to the WMR,
the Client Manager module monitors the set of clients
associated to wlanI and stores them in the Local Client
Table (LCTable).
The NDP-Proxy module provides the WMR with the
ability to act like a Neighbor Discovery Protocol proxy.
This allows to transparently forward packets toward the
WMR the destination client is associated with. In this
way, no particular mechanisms are necessary on the
client side, in order to communicate with remote clients.
The NDP-proxy module will correctly answer to the
ICMPv6 request sent by the local clients.
The IPv6 Forwarder module encapsulates all IPv6
packets in another IPv6 packet, in order to ship packets toward the destination client. This solution, though
introducing a certain amount of overhead, avoids keeping
state in the WMRs along the path between clients.
Indeed, only the WMRs at the edges of the path must be
aware that the two clients are communicating. Another
significant advantage is the robustness to mobility. If a
client changes the WMR to which it is associated, only
the WMRs at the edges of the path must update the
information (3 WMRs in total). WMRs along the path
do not need to make any update, while continuing to
relay packets.
In the next section we provide an example of how
communications are set up in a Wireless Mesh Network.
B. Communication Setup
Traffic flowing in a mesh network can be classified
into three types: a) traffic between two clients associated
Recall that in IPv6 broadcast addresses do not exist anymore,
multicast is used to send packets to multiple receivers.
with the same WMR, b) traffic between two clients
associated with different WMRs, and c) traffic between
a client and a gateway. All types of traffic are detailed
Clients associated with the same WMR. In this case,
clients do not need any particular attention, since the
wlanI interface automatically bridges the traffic between
them. This is an embedded feature of the HostAP mode
of wireless card drivers.
Clients associated with different WMRs. Let us take
the scenario of Figure 1 and suppose that client C1 wants
to send a packet to client C4 . The following steps show
how the communication between C1 and C4 is set up,
according to MeshDV rules:
1: C1 sends an ICMPv6 Neighbor Solicitation packet for C4 .
2: The NDP-Proxy module of WMR1 receives the packet from C1
and sends a message to the Client Manager module, requesting
to find out which WMR C4 is associated with.
3: The Client Manager module of WMR1 checks in the LCTable
and FCTable tables if the address of C4 is present. If it is not, it
sends a multicast packet (named MCREQ) on its wmrI interface,
in order to ask the other WMRs if someone knows C4 .
4: The Client Manager module of WMR3 receives the multicast
request and checks its own associated client list. It knows C4 ,
thus it sends a unicast reply (named CRREP). WMR3 keeps track,
in its local client table, that it exported the address of the client.
5: The Client Manager module of WMR1 receives the unicast reply.
It stores in its foreign table that C4 is associated to WMR3 and
sends a message to the NDP-Proxy module to announce that it
knows to which WMR the client is associated.
6: The NDP-Proxy module of WMR1 sends to C1 an ICMPv6
Neighbor Advertisement packet associating the IP address of C4
to the Ethernet address of the wlanI interface.
7: C1 receives the ICMPv6 packet and now sends the data packet
for C4 to WMR1 .
8: The data packet is captured by the Forwarder module of WMR1 ,
which encapsulates it in a packet destined to WMR3 and sends
9: The netBSD kernel of WMR3 receives the packet, decapsulates it
and automatically forwards it on the local wlanI interface where
C4 is associated.
10: C4 receives the data packet.
Figure 3 shows the temporal diagram for the above
mentioned exchange, which we call client advertisement
mechanism. Note that since the proposed mechanism is
symmetric, if C4 needs to send back a packet, the same
mechanism applies.
Client-gateway traffic. Traffic from a client toward a
gateway does not need any particular action, since the
Enhanced DV Module performs also Gateway advertisement. Traffic from the gateway toward a client follows
the same rules like traffic between clients associated to
different WMRs.
Client subnetwork
Local Client
Mesh subnetwork
Local WMR
Client subnetwork
Remote WMR
Remote Client
Fig. 3. Communication set up temporal diagram for clients associated to different WMRs.
In this section, we provide details about the data
structures of the Client Manager module, namely the
Local Client Table (LCTable) and the Foreign Client
Table (FCTable) and outline how they are managed.
Moreover, we present the set of packets necessary to
perform client advertisement, listing the format and the
purpose of each packet type.
A. Local Clients Table
The MeshDV routing demon is aware of the clients
that are associated to the WMR on which it is running.
For this purpose, it monitors the set of clients associated
with wlanI and stores them in the Local Client Table
(LCTable), which has the following format:
Associated ClientIp
Associated ClientIp
Associated ClientIp
Associated ClientIp
MAC address Client
MAC address Client
MAC address Client
MAC address Client
WMR_X Active
Each entry contains the IPv6 address of the associated
client, its MAC address, and a pointer to a list. The list
contains the IPv6 address and state information for all
the WMRs that have asked for that client, i.e., that have
sent a client request packet. Otherwise the content of
the field is NULL. State information is related to the
timeout mechanism, which is detailed in [12]. In the
above table, for example, the WMR has received from
WMR_X a request for Client1 and nothing for the rest.
MeshDV updates the table of associated clients by
accessing periodically the kernel, using ioctl calls.
After an update, some new clients may have associated
and some may have departed. For a newly associated
client, its IPv6 and MAC addresses are added to the
LCTable and the ExternWMRs field is initialized to
NULL. For a departed client, if ExternWMRs is not
NULL, for each WMRs in the list, a withdraw request
(CWIT) is sent. Once all WMRs in the list have replied,
the entry of the client is deleted from the table. An
LCTable update may also be triggered by the reception
of client request packets. In this case, the IPv6 address
of the request sender is added to the ExternWMRs list
of the entry corresponding to the requested client.
B. Foreign Clients Table
When a client wants to communicate to a client associated elsewhere, there is the need to know which other
WMR manages the destination client. MeshDV holds this
information in the Foreign Client Table (FCTable),
which has the following format:
0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
| Type
| Subtype
Requested Client IPv6 Address
Request Sender WMR IPv6 Address
Sender WMR CMM Request Sequence Number
Fig. 4.
MCREQ/UCREQ Packet format.
Foreign ClientIp
Foreign ClientIp
Foreign ClientIp
Foreign ClientIp
Manager WMR IP of Client
Manager WMR IP of Client
Manager WMR IP of Client
Manager WMR IP of Client
Each entry contains the IPv6 address for the foreign
client and the IPv6 address of the WMR with which
the foreign client is associated. The latter field is a list
that may contain several IPv6 addresses. Indeed, due to
transient phenomena caused by mobility of the nodes,
more than one WMR can send a CRREP packet to
advertise itself as manager for the same client. Except for
particular transient conditions, which, however, have limited lifetime, this field always contains only one WMR
address. Transient conditions are described in [12].
MeshDV updates this table upon receiving or sending
particular packets. When the WMR sends a MCREQ for
a foreign client, a new entry is added to the table. The
IPv6 address of the client is put in the ClientIP
field and the ManagerWMRIP field is initialized to
NULL. When the WMR receives a CRREP for the foreign
client, the IPv6 address of the reply sender is added
to the ManagerWMRIP list. When the WMR receives
a CWIT for a client, the IPv6 address of the packet
sender is deleted from the ManagerWMRIP list. If the
list becomes empty, the whole entry is deleted from the
C. Client Advertisement Packets
Here we present the format of the packets used to
perform client advertisement. For each packet we just
give its purpose, format, and the content of each field of
the packet. Their use is detailed in section V.
1) MCREQ: The Multicast Client REQuest (MCREQ)
packet is issued whenever a WMR needs to know which
WMR a client is associated with. As the name suggests,
this packet is always multicast. The format is presented
in Figure 4, and the content is:
Type: Set to 2, Client Advertisement Mechanism.
Unused: Sent as 0; ignored on reception.
Subtype: Set to 1, Multicast Client REQuest (MCREQ).
Requested Client IPv6 Address: The IPv6 address of the client
the WMR is looking for.
Request Sender WMR IPv6 Address: The IPv6 address of the
WMR who sent the request.
Sender WMR CMM Request Sequence Number: The request
Client Manager Module sequence number generated by the sender.
2) UCREQ: The Unicast Client REQuest (UCREQ)
packet is periodically issued in order to have reachability
confirmation, i.e., to confirm that the foreign client is still
associated to the same WMR. The packet format is the
same as for MCREQ, except for the subtype field which
is set to 2, and the fact that it is always unicast.
3) CRREP: The Client Request REPly (CRREP)
packet is issued when a WMR receives a MCREQ or a
UCREQ for a client that is in its LCTable. The WMR
sends back the CRREP as a unicast packet directly to
WMR that issued the request. The format is presented
in Figure 5, and the content is:
Type: Set to 2, Client Advertisement Mechanism.
Unused: Sent as 0; ignored on reception.
Subtype: Set to 3, Client Request REPly (CRREP).
Requested Client IPv6 Address: The IPv6 address of the client
the reply refers to.
Request Sender WMR IPv6 Address: The IPv6 address of the
WMR who sent the corresponding client request.
Client Manager WMR IPv6 Address: The IPv6 address of the
WMR who sent the reply.
Sender WMR CMM Request Sequence Number: The request
Client Manager Module sequence number generated by the sender.
4) CWIT: When there is a client that is not anymore
associated with a WMR, the corresponding entry in the
LCTable has to be deleted. A unicast Client WIThdraw
CWIT packet is sent to each WMR present in the
ExternWMRs list, in order to advertise the event. The
format is presented in Figure 5, and the content is:
0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
| Type
| Subtype
Requested Client IPv6 Address
Request Sender WMR IPv6 Address
Client Manager WMR IPv6 Address
Sender WMR CMM Request Sequence Number
Fig. 5.
CRREP/CWIT/CWREP Packet format.
Type: Set to 2, Client Advertisement Mechanism.
Unused: Sent as 0; ignored on reception.
Subtype: Set to 4, Client WIThdraw (CWIT).
Withdraw Client IPv6 Address: The IPv6 address of the client
the withdraw request refers to.
Withdraw Sender WMR IPv6 Address: The IPv6 address of
the WMR who sent the withdraw request.
Withdraw Destination WMR IPv6 Address: The IPv6 address
of the WMR the request is addressed to.
Sender WMR CMM Sequence Number: The request Client
Manager Module sequence number generated by the sender.
5) CWREP: The Client Withdraw REPly (CWREP)
packet is issued when a WMR receives a CWIT for a
client that is in its FCTable. The WMR updates its
FCTable and sends back the CWREP as a unicast packet
directly to the WMR that issued the request. The format
is presented in Figure 5, and the content is:
Type: Set to 2, Client Advertisement Mechanism.
Unused: Sent as 0; ignored on reception.
Subtype: Set to 5, Client Withdraw REPly (CWREP).
Withdraw Client IPv6 Address: The IPv6 address of the client
the withdraw reply refers to.
Withdraw Sender WMR IPv6 Address: The IPv6 address of
the WMR who sent the corresponding withdraw request.
Withdraw Destination WMR IPv6 Address: The IPv6 address
of the WMR who sent the reply.
Sender WMR CMM Sequence Number: The request Client
Manager Module sequence number generated by the sender.
As mentioned in section II-A, the MeshDV demon acts
also as NDP-proxy on the wlanI interface, by means of
a specific module. We detail here this module, before the
Client Manager algorithm, in order to give a complete
overview of the events that the latter has to handle.
To perform its task, the NDP-proxy module needs to
capture ICMPv6 packets, opening an ICMPv6 socket.
Algorithm 1 NDP Proxy Algorithm
until (New Event Occurs)
if ( New Event == Neighbor Solicitation Packet Received ) then
TargetClientIP = IPv6 address of requested client
Ask the Client Manager Module if TargetClientIP is in LCTable
if TargetClientIP not in LCTable then
if (Neighbor Solicitation was sent in multicast) then
Trigger Client Manager module to send a MCREQ
else if (Neighbor Solicitation was sent in unicast) then
Trigger Client Manager module to send a UCREQ
end if
end if
else if ( New Event == Client Manager Module Message Received) then
CREQResult = Client Manager Module Message State
if (CREQResult == Client Found) then
Send a Neighbor Advertisement packet on wlanI
end if
end if
end loop
It filters out all packets except Neighbor Solicitation
requests, which need to be exanimed. Details on how to
use ICMPv6 sockets and how to filter the messages are in
[7], while details about ICMPv6 protocol and messages
are in [9] and [8].
The NDP-proxy module waits for an event to occur.
An event can be the reception of a Neighbor Solicitation
packet, through the ICMPv6 socket, or a message from
the Client Manager module. If the event is a Neighbor
Solicitation packet, the IPv6 address of the target client
is extracted. If the target client is in the LCTable, the
client itself will reply directly since it is in the local
WLAN; no further action is taken. Otherwise, the NDP
module triggers the Client Manager module to send a
client request packet. If the Neighbor Solicitation packet
was a multicast packet, a MCREQ request is triggered.
Otherwise, if it was a unicast packet, a UCREQ request
is triggered.
The Client Manager module manages request timeouts
and always provides anyway an answer to the NDP
module. If the Client Manager module answers that a
reply is received, the NDP-proxy sends back a Neighbor Advertisement packet. This packet sets the MAC
address of the wlanI interface of the WMR as the
MAC address of the target client. With these settings,
future data packets, sent to the target client, will pass
through the wlanI interface, captured by the Forwarder
module, encapsulated, and correctly routed inside the
mesh network.
The complete algorithm is described via a pseudo language in Algorithm 1. Note that the NDP-proxy module
does not implement an active wait; in the algorithm this
action is defined as an active loop only for the sake of
The Client Manager module has two main tasks:
the consistent management of both LCTable and
FCTable structures and the client advertisement communication on the mesh subnetwork, through the wmrI
interface. The module is also responsible for maintaining
a local counter named CMM sequence number, i.e., the
Client Manager Module sequence number. This counter
is used whenever MCREQ/UCREQ/CWIT packets are issued. The counter is increased by one each time a
request is sent. The CMM sequence number increases
robustness, since for all requests, the corresponding reply
must carry the same sequence number in order to be
accepted. Moreover, it limits broadcast overhead. The
MCREQ is a multicast packet, sent to FF02::2 (All
link-local IPv6 Routers). All one-hop away WMRs will
receive the request and will forward the packet if they
are not able to reply to the request. In order to forward
an MCREQ request only once, each WMR keeps a perWMR CMM sequence number, stored in the routing
table maintained by the Enhanced DV module. The
MCREQ is forwarded only if the requested client is not
associated with it and the carried CMM sequence number
is higher than the one already stored.
Note that the CMM sequence number is not the same
sequence number as described by Perkins et al. in [1].
The sequence number proposed in [1] is associated
with each destination, in order to avoid loops, and is
embedded in each entry advertised by route update messages. This feature remains unchanged. The new CMM
sequence number is only used in Client Advertisement
packets. Route update messages do not carry the CMM
sequence number.
When a WMR receives a request message, it updates
the CMM sequence number value and processes the
message only if the new CMM is higher than the old
value. Otherwise the packet is silently discarded.
The Client Manager module reacts only to four types
of events: 1) a request has been received; 2) the NDPproxy module has made a request; 3) a timeout has
occurred; 4) a client has associated/disassociated to the
wlanI interface.
In the following subsections, we describe how the
Client Manager module works. The main loop is not
described since it is simply an event handler that calls the
right subroutine. Furthermore, since the CWIT sending
subroutine consists of just preparing and sending the
packet, it is not detailed. For the sake of clarity and
simplicity, timeout events are not described; a detailed
description can be found in [12].
A. Packet reception subroutines
1) MCREQ reception subroutine: When a MCREQ is
received, the WMR first checks whether the sequence
number is already stored in the routing table. If so, no
Algorithm 2 MCREQ Reception Subroutine
if (CMM Request Sequence Number higher than the one stored in RTable) then
Update the CMM Request Sequence Number in the correct entry of the RTable
if (Request Client in LCTable) then
if (Request Sender not in ExternWMRs list) then
Add Request Sender in the corresponding ExternWMRs list
end if
Prepare CRREP Packet
Send unicast CRREP Packet to Request Sender
Forward MCREQ Packet on wmrI using FF02::2 address
end if
end if
Algorithm 3 UCREQ Reception Subroutine
if (CMM Request Sequence Number higher than the one stored in RTable) then
Update the CMM Request Sequence Number in the correct entry of the RTable
if (Request Client in LCTable) then
if (Request Sender not in ExternWMRs list) then
Add Request Sender in the corresponding ExternWMRs list
end if
Prepare CRREP Packet
Send unicast CRREP Packet to Request Sender
end if
end if
further action is needed since the received MCREQ is a
duplicate. If not, the sequence number is stored and the
subroutine checks if the request should be forwarded
or replied to. This last decision depends on whether
the requested client is associated to its wlanI or not.
Since any MCREQ retransmission increases the sequence
number, if a MCREQ, whose sequence number is already
stored, is received, it is considered as a duplicate packet.
The subroutine is detailed in Algorithm 2.
2) UCREQ reception subroutine: When a UCREQ is
received, the WMR first checks whether the sequence
number is already stored. If so, no further action is
needed since the received UCREQ is stale or duplicate. If
not, the sequence number is stored, and the subroutine
checks if the request needs to be replied to. Again,
this action depends on whether the requested client is
associated to its wlanI or not. The subroutine is detailed
in Algorithm 3.
3) CRREP reception subroutine: When a reply with a
correct sequence number is received, the sender is added
in the correct entry of the FCTable list and the NDPProxy module is informed. The subroutine is detailed in
Algorithm 4.
4) CWIT reception subroutine: When a CWIT request
is received, the CMM sequence number is checked to
see if the request is a new one. If so, it does not
matter if the client has been already withdrawn from the
FCTable or not, the WMR replies regardless. Indeed, if
a WMR receives a new withdraw request for an already
withdrawn client, this means that the previous reply got
lost. The subroutine is detailed in Algorithm 5.
5) CWREP reception subroutine: When a withdraw
reply with a correct sequence number is received, the
sender is deleted from the ExternWMRs list of the
Algorithm 4 CRREP Reception Subroutine
if (Request Sequence Number has no match in Pending Requests Queue) then
Discard Packet
Stop corresponding timeout timer
ClientIP = Target Client IP of the request
Delete corresponding request packet from Pending Requests Queue
if (Client Manager WMR not in FCTable for ClientIP) then
Add Client Manager WMR to ManagerWMRIP list for ClientIP
end if
Message to NDP-Proxy module ClientIP reply received
end if
Algorithm 6 CWREP Reception Subroutine
if (Request Sequence Number has no match in Pending Requests Queue) then
Discard Packet
Stop corresponding timeout timer
ClientIP = Target Client IP of the request
Delete corresponding request from Pending Requests Queue
delete Client Manager WMR from ExternWMRs list for ClientIP
end if
Algorithm 7 MCREQ Send Subroutine
Algorithm 5 CWIT Reception subroutine
if (CMM Request Sequence Number higher than the one stored in RTable) then
Update the CMM Request Sequence Number in the correct entry of the RTable;
Prepare CWREP Packet;
if (Withdraw Client is in FCTable) then
Delete Withdraw Sender from ManagerWMRIP list;
if (ManagerWMRIP list is empty) then
Delete Withdraw Client entry from FCTable;
end if
end if
Send CWREP Packet
end if
correct entry of LCTable. The subroutine is detailed
in Algorithm 6.
B. NDP-proxy requests subroutines
1) MCREQ send subroutine: This subroutine simply
prepares the packet and sends it to the multicast address
FF02::2, on the wmrI interface. When the packet is
created, the CMM sequence number is increased by 1.
The subroutine is detailed in Algorithm 7.
2) UCREQ send subroutine: This subroutine is very
similar to the MCREQ send subroutine, the algorithm is
not depicted here. The main difference is that the packet
is unicast to the client manager indicated in the entry of
the FCTable.
C. Client Disassociation subroutines
1) CWIT send subroutine: This subroutine is very
similar to the UCREQ send subroutines, except that the
destination is taken from the LCTable. The algorithm is
not detailed since it just consists of preparing the packet
and sending it on the correct interface.
In order to avoid maintaining state information along
the WMR-path connecting two clients associated to a
different WMR, the MeshDV uses a tunneling approach.
When a client sends a packet to a remote client, because
of the NDP-proxy module, the packet is delivered to
the local WMR. The local WMR is now in charge
of forwarding the packet. At the IP level, the packet
contains the local client address as a Source Address and
the remote client address as a Destination address. The
kernel is not able to route this packet because it is not
aware of remote clients. Nevertheless, the MeshDV has
if (Requested Client IP is not in FCTable) then
Create a new entry in FCTable for the requested Client IP
end if
Prepare MCREQ Packet
Send MCREQ Packet multicast FF02::2
Append MCREQ to Pending Requests Queue
Set timeout counter at CRREP_TIMEOUT
Start timeout Timer
all the needed knowledge and can easily create a new
IPv6 packet destined to the remote WMR, containing
the original one. Since routes between WMRs are setup
proactively, WMRs along the path are able to route the
new packet without even knowing the destination client
address. Only the WMRs where the two clients are associated are aware of the ongoing communication. Indeed,
they have the local client and the remote client addresses
respectively in the LCTable and the FCTable.
An alternative solution, in order to route a packet
toward the WMR with which a client is associated, is
the Routing Header option of IPv6. Nevertheless, since
WMRs along the path do not know the source client, in
case of packet drop, they would try to send an ICMP
packet to the source client. But the kernel is unaware
where the client is and unable to find it since it has
access only to the local wlanI. MeshDV is able to find the
client but it should be aware of all packets dropped by the
kernel, turning into a very complex software architecture.
Using tunneling makes MeshDV simpler. In the following,
the operations of encapsulation and decapsulation are
Decapsulation. Tunneling needs to be managed by the
MeshDV demon only when the packet enters the mesh
network, i.e., when it comes from the wlanI interface and
is forwarded on the wmrI interface. In the other direction,
when a tunneled packet arrives at the destination WMR,
the netBSD kernel is able to decapsulate it. Then, since
the kernel is aware of the clients associated to the wlanI
interface, it is able to deliver the inner packet without
any additional operation.
Encapsulation. The encapsulation is done inside
MeshDV, because the kernel is not aware of which
WMR the destination client is associated with. All IPv6
packets destined to remote clients that arrive at the wlanI
interface are directly delivered to the Forwarder module,
in order to be encapsulated. They must not pass through
the kernel. For this purpose, on wlanI, PF [10] is used to
filter them; they are captured at link-level and delivered
to MeshDV by the Pcap [11] utility. The packet received
from the Pcap filter is extracted from the layer 2 frame, a
new IPv6 header is pre-pended to it, and finally delivered
to the kernel. The kernel is now able to route the packet,
through the wmrI interface to the correct WMR.
The testbed on which the MeshDV is currently being
implemented is composed of custom WMRs assembled
in our laboratory. Each WMR is a Soekris net4521 box,
running NetBSD 2.0, and using only the IPv6 protocol
stack. The wlanI interface is a Proxim 8470-WD b/g card
set in HostAP mode. The wmrI interface is a NetGate
5354 MP Plus Aries2 4G a/b/g set in ad hoc mode, using
MeshDV implementation is not yet completed at the
time of this writing, nevertheless, some initial observation can be underlined. The first observation we achieved
while implementing MeshDV is that global unique addresses are useless on the wmrI interface. Indeed, it is
possible to reach any WMR by simply using the wlanI
global address by correctly setting the kernel routing
table. Interfaces of WMRs in the mesh subnetwork have
a Link-Local unicast address, which is not unique and
has a local scope. Thus, taking the topology of Figure 1,
the kernel routing table of WMR2 concerning route to
WMR4 and WMR1 becomes:3
MeshDV also performs gateway advertisement. Thus, a
default route is always present on each WMR, as shown
in the above example. This allows to sligthly reduce
overhead, since no special messages need to be issued.
Kernel routing table size is always proportional to the
number of WMRs. MeshDV routing table size is given
by the sum of the size of the routing table inside the
Enhanced DV module (proportional to the number of
WMRs) and the size of the LCTable and FCTable
tables. The sizes of these tables are proportional to the
number of clients present in the network and the ongoing
communication between clients associated to different
WMR. The FCTable of the gateway WMR contains
In our
client subnetwork
2001:660:3302:2A21/64 at which the WMR number is added in
order to have the global address of the wlanI interface. Furthermore,
the MAC address of WMR4 wmrI interface is 00:02:6F:20:F7:ED.
the list of all clients of the network. Indeed some basic
services, like DNS, are placed outside the mesh, thus,
there is always traffic from and to all the clients. In [12]
we present in detail the reduction in the routing table
size that MeshDV can achieve.
Concerning scalability, we expect MeshDV to scale
like RIPng [6], since the internal routing paradigm is
basically the same.
In this paper, we presented the prototyping research
work for the implementation of MeshDV, a routing architecture for wireless mesh networks. We described the
general functionality of MeshDV and detailed its main
components. At present, tests are being performed on a
testbed deployed at the LIP6 laboratory, aiming to deploy
a large WMN covering the campus of the University of
Paris VI.
Besides the fact that the protocol is implemented over
IPv6, there are other features that make it interesting.
MeshDV maintains routes between WMRs in a proactive
fashion, while searching for routes toward clients in an
on-demand fashion. The architectural characteristics of
MeshDV have been strategically chosen in order to obtain
the following advantages: 1) reduced routing table sizes,
since there is no need for keeping all routes toward all
clients; 2) easy management of clients that change WMR
association, since related information is maintained only
on edge WMRs.
[1] C.E. Perkins, P.R. Bhagwat, “Highly dynamic destinationsequenced distance vector routing (DSDV) for mobile computers,” in Proceedings of ACM SIGCOMM, London, sep. 1994.
[2] B. Gupta, “Design, Implementation and Testing of Routing Protocols for Mobile Ad-Hoc Networks”, Master Thesis,
University of Illinois at Urbana-Champaign, 2002 Online at: wireless/dsdv/
[3] “LocustWorld Bio Diverse Networking Unleashed!”,
Online at:
[4] “MIT Roofnet”,
Online at:
[5] “RFC 2460: Internet Protocol, version 6 (IPv6) Specification”,
[6] “RFC 2080: RIPng for IPv6”, 1997.
[7] “RFC 2292: Advanced Socket API for IPv6”, 1998.
[8] “RFC 2462: Neighbor Discovery for IP Version 6 (IPv6)”, 1998.
[9] “RFC 2463: Internet Control Message Protocol (ICMPv6) for
the Internet Protocol Version 6 (IPv6) Specification”, 1998.
[10] “PF: The OpenBSD Packet Filter”,
Online at:
[11] “Programming with Pcap”,
Online at:
[12] L. Iannone, “MeshDV: Implementation Draft”,
Online at:˜iannone/meshDV.pdf
Posters and Demos
Social networks, novel communication applications and needs in mobile
Claudia Brazzola
University of Applied Sciences of Lugano, Dept. of Social and Administrative Sciences
[email protected]
This paper deals with how some communication
technologies and applications intervene in modeling
and changing the characteristics of our modern society
structure. It starts from the concepts of mobility, of
strong and weak ties and concludes by analyzing some
examples of novel communication applications with
focus on applications for mobile ad hoc networks and
by presenting some results from studies about the
elderly and new technologies of communication,
conducted within the MobileMAN project.
1. Introduction
The aim of this paper is to reflect on
communication applications, with particular interest in
mobile applications of ad hoc networks and the
implications that this has on the structure of our
modern society. Some of the content is the result of
studies conducted within the MobileMAN project that
had the focus on the potential consequences of ad hoc
networks on users. The MobileMAN project is a EU
project of the Information Society Technologies area
that aims to design, implement and validate a new
paradigm of ad hoc networks, from the technical,
economic and social points of view.
2. Modern Social Networks
One’s social network is the complex of
relationships with other individuals. Strong ties are
those such as kin relations and close personal friends,
whereas weak ties are loose acquaintances such as
those connections made at a party. Following
Granovetter’s thought [1] we consider both types of
ties as vital for an individual and analyze in what way
technology and existing or future applications can help
maintain and build both weak and strong ties.
3. Mobility of Individuals – New Needs and
New Opportunities
3.1. Internet + Mobility = ?
We start from two central concepts that we are
identifiable in our society: the increasing importance of
the Internet connectivity and the high mobility of
individuals. Internet has been changing the way we
live our life. The next element is mobility: numerous
workers are commuters and many people are rarely at
home. Portability of communication means that
individuals are virtually able to communicate and
connect to networks anywhere and anytime. They must
not wait to be in a particular place [2]. Putting together
these two aspects, a whole range of needs result. In
one of his articles, Rheingold writes “[...] participants
in online communities will remain in continuous
contact over multiple platforms on desktops and in
mobile devices, and will be used to coordinate group
activities in the geographic world, thus blending
affinity-based and local-acquaintance-based social
communication.” [3] He speaks of the combination of
the Internet world and the mobile world, with which
the user will interact continuously.
3.2. Implications for MobileMAN – New
Applications / Scenarios: an Experiment with a
Wiki Website
Mobile phones allow people to communicate in
many ways such as SMS, MMS, e-mail. However,
they do not allow expanding social networks – they
only help individuals in managing existing contacts. In
fact, “mobile communications are organized around
known social networks. People call and message
people they already know. Most often, you
communicate with people who are already in your
address book.” [4]. This is also supported by data from
a study on mobile phones and ad hoc networks
currently ongoing at SUPSI. So far, participants in this
qualitative study agree on saying that mobile phones
do not help expanding social networks, but
maintaining and developing existing relationships.
We believe that there is an untapped market for
applications that allow expanding social networks,
more than those applications that enable individuals to
manage existing relationships. MobileMAN could
exploit this opportunity by developing applications that
are oriented to this opportunity. An experiment of
MobileMAN project and involving a number of
students of Helsinki University of Technology aimed
at developing new applications and scenarios for ad
hoc networks through group collaboration either using
offline medium (paper) or online medium (wiki
website [5]). A wiki is a particular type of site whose
content is quickly editable through any browser
without the need for the user to have any programming
skill. The experiment had also the objective to verify
the appropriateness of wiki as a tool for participatory
design activities in technological and innovative fields.
Students created some scenarios of use for
MobileMAN in small groups, and in a second round
were asked to complete and comment on any three
chosen scenarios among the developed. This exercise
was within one course of their curriculum in
networking and was not compulsory; who participated
received a reward for the course final grade.
Participation was not too extensive – 27 students (of a
class of 112) worked in the scenario developing
activity, and only 20 collaborated until the end of the
exercise by also filling in a questionnaire, which
covered various topics, from an evaluation of the
exercise itself to opinions about the future
development of ad hoc and infrastructure-based
networks. We interpreted this low participation with
the abstractness of the concept of ad hoc. The exercise,
in fact, revealed also the difficulty to imagine
applications that are exclusive for ad hoc networking
devices, most likely because of the abstract nature of
the ad hoc concept. Being based on cooperation of the
users, it relies on a completely new model of low-level
architecture. Applications that run on top can be the
same as those that run on mobile phones or others. The
hypothesis is that innovative applications that
empower the user to expand their social network will
be more interesting and appealing. Ad hoc networks
are infrastructure-less and consequently (ideally) with
no use cost for the user. This, however, cannot be the
only driving motive that pushes for adoption –
especially if the quality of service is not as high as is
now the quality of mobile phones applications (voice
clarity, network availability). In this vision,
applications can play a fundamental role in the
adoption process of the technology. We present the
summarized content of two scenarios that are
particularly valuable, developed by two groups of
Public festival scenario – a public festival is an
occasion where there are many people gathered in a
limited area and it can happen that base stations are
stuck because of overload. In such situation ad hoc
would be helpful. In the developed scenario, three
boys use their MobileMAN device to micro coordinate
themselves and to receive real-time information about
facilities. Another application in this scenario is a
profile manager that alerts the user if there is in close
vicinity some user that matches their profile. This is an
application designed for extending social networks.
Although some aspects of this scenario need further
analysis of realistic realization it provides interesting
ideas to work on.
Where to party scenario – big cities often offer
far too many options to spend free time and it may be
difficult to make a choice. This application would
combine location information, by informing of
opportunities in the user nearby, social networking
expansion, by viewing the structure of the network,
with for example, friends of their friends, and the
digital equivalent of “word of mouth”, that is,
MobileMAN users post information and broadcast it to
the network of persons they are connected to.
SMS has been the killer application for mobile
phones with millions of messages sent every day.
However, chat-SMS1 have not so far had great success
because “SMS connects people very inefficiently.
Those who design future services would do well to
search for more efficient ways of connecting people.”
[6] The last example of scenario could be a response to
this need for more efficient applications of sending
3.3. MobileMAN Whiteboard
Within the MobileMAN project, a peer-to-peer
multicast application called whiteboard is being
developed. It can be considered as a first version of a
distributed chat platform for text writing or graphic
drawing. Comparing this application with traditional
SMS written and sent by mobile phone, we can say
that it is certainly interesting, since it places in between
SMS and chat. Being a multicast system, it would
Chat-SMS is a system where users subscribe to a chat service
(channel) and receive each message sent by all other participants to
the channel. The dubious issue in this service is that “it remains to be
seen how willing the participants in the chat groups are to pay for
EVERY message sent to the chat channel”. [7]
increase the limited efficiency of SMS – something
that is in line with Saarikoski’s thought [6].
3.4. The Elderly – the “Grey Digital Divide”
The elderly constitute a particular target when
defining applications and services for specific groups:
they have a range of needs that at the moment are
unaddressed. MobileMAN expressed the intention to
privilege this group of individuals and potential end
users of MobileMAN, based on the belief that
technology can help elderly people improving their
life. Therefore, a study of their relationship with new
ICTs is currently being carried out at SUPSI within the
project. The study focuses on technologies like TV,
VCR, DVD, computer and internet, mobile phone and
telecare wristalarm2. All participants had a telecare
wristalarm installed and were interviewed about their
personal and social situation, existing relationships and
activities and technology they use in their everyday
life. 10 out of 11 interviewed lived alone (which was
the reason why they adopted the telecare wristalarm).
The initial idea was that for elderly people living
alone, new ICTs would have been particularly valuable
as they would have allowed a higher communication
and interaction with society, even if physically
impaired (difficulty to go out and walk e.g.). However,
findings demonstrated that the elderly have a very low
interest in new ICTs, even if these can help them
overcoming their isolation, which is – as said – very
often a characteristic of their social and personal
situation. As a consequence, it is very difficult to
involve this category of people in participatory design
activities within MobileMAN. We explained this
disinterest towards innovative communication
technologies (mobile phone, internet, chat) as related
to their particular life path: they lived through war
time, and lived generally without all the
communication means we dispose of nowadays. This
belonging to a different world is very clear in the fact
that almost all the interviewed elderly expressed the
inability to understand the need for communication and
therefore considered as useless communication
empowering technologies (chat, email, internet in
general, mobile phones). This result confirms findings
of Millward [8], who researched the issue of “grey
digital divide”. This way of thinking about new ICTs,
however, is probably only typical of this generation of
This is a system created for elderly who live alone and are at risk of
falling or have health problems. It is wrist button that can be pushed
in case of necessity and that contacts directly the emergency station,
which deals with the situation. This system has proven very useful
since there is no time loss due to the inability of the person to reach
the phone in case of emergency.
elderly: in 15-20 years the old people will have a
different relationship and opinion about them. The
“grey digital divide” might diminish naturally with
4. Conclusions
In this paper we have started from the networked
structure of society. We presented two scenarios
developed by two groups of students within an activity
of the MobileMAN project, one application
(whiteboard) currently being implemented by
MobileMAN partners that can provide some
interesting ideas for novel applications to be developed
for ad hoc network devices. Moreover, we presented
some results of a study conducted with a number of
elderly people on their relationship with new
communication technologies. The main results confirm
the existence of a “grey digital divide” that might
however diminish with time.
5. References
[1] M. Granovetter “The Strength of Weak Ties: A
Network Theory Revisited”, Sociological Theory
Vol.1, New York, 1983, pp. 201-233
[2] B. Wellmann, “Physical Place and Cyberplace: The
Rise of Personalized Networking”, International
Journal of Urban and Regional Research 25, Special
issue on “Networks, Class and Place” ed. Talja
Blokland and Mike Savage, Toronto, 2001
[3] H. Rheingold, “Mobile Virtual Communities”,
2003, published on
[4] H. Rheingold, op.cit.
[6] Saarikoski quoted in H. Rheingold, “Email, ScaleFree Networks, and the Mobile Interne”, 2005
[8] P. Millward, “The ‘grey digital divide’: Perception,
exclusion and barriers of access to the Internet for
6. Acknowledgments
Many thanks to Prof. Dr. Christian Marazzi and
Dr. Jennifer Duyne for reading this article and for their
precious comments and suggestions.
On the Dimensionality of Wireless Connectivity Traces
George Roussos
Birkbeck College
University of London
[email protected]
Routing efficiency in wireless networks can be
greatly improved by matching mobile host connectivity
patterns. To this end, over the past few years considerable effort has been invested in developing predictors of mobility patterns that is, models of mobile host
movement so that for a specific sequence of recorded
locations to predict the most likely subsequent location. In this paper, I initiate a study of the structure of
the connectivity space itself. I analyze samples from
the Dartmouth data set1 and conduct an exploratory
study of the structure of the underlying space. In
particular, I compute the singular values of the timeweighted connectivity matrix, and relate this result to
principal component analysis. Initial findings indicate
that the degrees of freedom of the space induced by the
connectivity matrix is very low with respect to the number of access points and mobile hosts involved; that the
dimensionality of the underlying space grows slowly
with the size of the sample; and, that the distribution
of its eigenfrequencies follows a power law.
1. Introduction
Wireless networks can better serve mobile hosts by
employing client location information to better anticipate connectivity patterns. Predicting accurately the
location of hosts can potentially improve the performance of wireless routing and the robustness of the
network infrastructure itself, thus improving the user
experience for a variety of applications. These improvements lead to a better user experience, to a more
cost-effective infrastructure, or both. As a result, during the past few years a number of location predictors
have been proposed in the literature based on a variety of complementary techniques including Markovbased, compression-based, PPM, and SPM mechanisms. Such approaches infer models of mobile host
movement patterns so that for a specific recorded sequence of recorded locations to predict the most likely
subsequent location. A comprehensive comparative
evaluation study of the relative performance of several
of these algorithms on the Dartmouth mobile data trace
[2] including a detailed description of their structure
and performance can be found in [4].
In this poster, rather than focusing on predictors I
initiate the study of the structure of the connectivity
space itself so as to understand its core characteristics.
To do this, I also employ the Dartmouth data trace to
reconstruct the time-weighted connectivity matrix between access points and mobile clients which we use
as the basis of this investigation. Several techniques
are employed to explore the structure of this space,
with a view in all cases to identify the existence of
a small subset of components or eigenfrequencies that
characterize accurately the client connectivity behavior. Effectively, I aim to determine a basis for a lowdimensional projection of the connectivity matrix that
provides an appropriately accurate approximation to
the overall connectivity patterns.
2. SVD and Principal Component Analysis
Many thanks too David Kotz and other members of the Dartmouth Centre for Mobile Computing for providing access to this
Let L = {lij } be the m × n connectivity matrix
of a data trace defined by setting the element lij to be
proportional to the time mobile host j is connected to
access point i, for a data sample that includes traces
of n hosts and m base stations. Note that row i of L
describes the time each sighted client has spend connected to base station i, and column j of L describes
the time spend by host j connected to each base station. Hence, I refer to the row vectors of L as the connectivity profile for access point i and to the column
vectors as the connectivity pattern of client j. Due to
the rather limited mobility of hosts in the Dartmouth
traces, the matrix L is sparse.
The singular value decomposition (SVD) of L is
L = U SV T ,
where U is an m×n matrix, S is an n×n diagonal matrix, and V also an n×n matrix. The columns of U are
called the left singular vectors and form an orthonormal basis for the connectivity patterns of clients, that
is ui uj is zero except where i = j when it is one. The
rows of V T contain the elements of the right singular
vectors and form an orthonormal basis for the connectivity profiles for the access points.
The matrix S is zero everywhere except at the diagonal, that is S = diag(s1 , . . . , sn ), where it contains
the so-called singular values sk of L. Singular values are ordered so that the highest singular value is in
the upper left index of the matrix S. Note that if L is
square, that is m = n, then the SVD is equivalent to
the solution of the eigenvalue problem.
With the SVD at hand, we can compute the closest
r-rank matrix to L as follows
L(r) =
uk sk vkT
so that L(r) minimizes the sum of the squares of the
difference of the elements of L and L(r) . Standard
approaches to compute the SVD can be found in [3].
Relation to principal component analysis
Principal component analysis (PCA) captures the
variance in a dataset in terms of its so-called principle
components. The SVD is intimately related to PCA
when principal components are calculated using the
covariance matrix. If each column of L is centered
then LT L is proportional to the covariance matrix of
the connectivity profile for the access points. Moreover, diagonalisation of LT L yields V T and thus the
principal components of the connectivity profile. in
other words, the right singular vectors of L are the
same as the required principal components. Further,
the eigenvalues of LT L are the singular values of L,
which are proportional to the variances of the principal
components. Overall, the matrix U S contains the principal component scores, which are the coordinates of
the connectivity profile in the space of principal components. If instead of centering the columns of L we
center its rows then LT L is proportional to the covariance matrix of the connectivity patterns of the mobile
clients. Similar to above, the left singular vectors are
also the principal components of the connectivity pattern space; the singular values are proportional to the
variances of the principal components; and the matrix SV T contains the principal component scores, that
is the coordinates of the connectivity patterns in the
space of principal components.
In this section I report on preliminary results of the
analysis of connectivity patterns using the Dartmouth
mobile data trace. Logfiles provided by the CMC were
post-processed to extract the connection/disconnection
patterns of particular hosts to particular base stations.
Several data sets where developed consisting of up to
90, 000 samples, which were subsequently analyzed
following the discussion in previous sections.
As noted earlier the main focus of the analysis is to
better understand the underlying structure of this space
which in this case is characterized by the range of the
singular values. For the largest sample we find that
the majority of the singular values are relatively small
with respect to the largest components, in fact only 4
of those are within 10% of the magnitude of the largest
one. Figure 1 provides some more information regarding the actual distribution of the spectrum: the magnitude of the computed singular values appears to follow
a power law distribution.
Both observations point towards the fact that it
should be possible to reconstruct the full connectivity
matrix using only a small number of principal components within a high accuracy. Indeed, using only the
29 principal components it is possible to achieve accu-
racy of the order of 0.1% for this sample. Thus, the
variability of the connectivity patterns in the data set is
very low and can be predicted very well using a subspace of very low dimensionality.
Loglog plot of singular values of connectivity matrix
Figure 1. Logarithmic plot of the singular values of the connectivity matrix by size (90, 000
3. Discussion and Conclusions
Several papers on wireless networking measurement observe that mobility patterns of individual hosts
are likely to contain a considerable amount of periodicity. The underlying reasons for this are mainly economic and demographic, and dictate that clients move
within a small range of different velocities, and travel
along similar routes with most journeys starting and
ending at similar places. As a consequence, there is
only a small number of approximately discrete frequencies characterizing the behavior and the connectivity profile of the mobile station (and its user). In
this paper I provide some preliminary results regarding
the dimensionality of the space induced by the mobility patterns observed within a particular experimental
setting. Moreover, some early results are highlighted
regarding the number of degrees of freedom that exist
in the data and have estimated the order of accuracy
of reconstruction using a low-dimensional approximation.
This evidence can be used to potentially improve
the performance of wireless routing and the robustness
of the wireless network infrastructures. For example,
using the computed principal components it is possible
to pinpoint bottlenecks in wireless networks as well as
loss points for example due to interference. Knowledge of such locations allows to locate additional resources where they are needed to improve reliability
for example hop-by-hop rather than end-to-end packet
loss recovery.
More importantly, these findings provide evidence
in support of a recent conjecture by Jon Crowcroft [1]
regarding the feasibility of a common network architecture that overarches both mobile wireless mesh and
fixed networks. He observes the following correspondence between wireless mesh and wireline fixed networks:
Mesh network
Wireline network
Freq. distribution
Router out degree
A critical element for the proposed unified reference
model is that the distribution of journey frequencies
follows a similar pattern to the distribution of communication popularity, as recorded in the out degree
of the connectivity graph of Internet routers for example. This would imply that the routing system for both
types of networks would have similar properties.
In this paper I provide a strong indication that this is
indeed the case, since the computed singular values of
the connectivity matrix or else the journey frequencies
clearly appear to follow a power law distribution (cf.
Figure 1) and is thus qualitatively similar to the router
out degree on the Internet.
[1] J. Crowcroft. Communication to the UCL Mobile Systems Group. 10 September 2004.
[2] D. Kotz and K. Essien. “Analysis of a Campuswide Wireless Network”. Wireless Networks,
11:115-133, 2005.
[3] W. H. Press, B. P. Flannery, S. A. Teukolsky, W.
T. Vetterling. Numerical Recipes in C: The Art
of Scientific Computing. Cambridge University
Press, 1992.
[4] L. Song, D. Kotz, R. Jain, X. He. “Evaluating
next-cell predictors with extensive Wi-Fi mobility data”. Proc. 23rd Annual Joint Conference of
the IEEE Computer and Communications Societies (INFOCOM):1414-1424, 2004.
Cross-Layer Support for Group-Communication Applications in MANETs∗
Marco Conti, Franca Delmastro
Jon Crowcroft, Andrea Passarella
CNR, IIT Institute
Via G. Moruzzi, 1 – 56124 Pisa, Italy
University of Cambridge, The Computer Laboratory
15 JJ Thomson Avenue – Cambridge CB3 0FD, UK
P2P systems are a natural way of supporting groupcommunication applications in MANETs. In this paper we
discuss our experiences in developing such an application
in the real world. We highlight limitations of legacy P2P
systems, and show that solutions based on cross-layer optimisations are very promising.
1 Introduction
One of the most interesting class of applications that
can be envisaged for MANETs is represented by groupcommunication applications. In the framework of the MobileMAN Project [6], we are investigating the viability of
developing such kind of applications on real ad hoc networks. To this end, we developed the Whiteboard application (WB), which implements a distributed whiteboard
among MANET users. WB usage is very intuitive (see Figure 1). Each MANET user runs a WB instance on her device, selects a topic she wants to join, and starts drawing
on the canvas. Drawings are distributed to all nodes subscribed to that topic, and rendered on each canvas. We believe that these simple, “Plug&Play” applications will be of
great value for MANET users.
Developing this kind of applications in MANETs is a
challenging task. In this paper we present the networking
solutions we have studied and tested to this end. We present
alternative networking frameworks for supporting WB-like
applications (Section 2). Then, we compare a standard P2P
system (Pastry [8]) with CrossROAD [5], the P2P system
optimised for MANETs that we have designed within these
frameworks (Section 3). Advantages of the CrossROAD approach are presented by means of experimental results in
Section 4. Finally, Section 5 concludes the paper.
2 WB integration in MANETs
Group-communication applications such as WB are distributed, self-organising, decentralised in nature. Designing
them on top of P2P systems guarantees a great flexibility
and optimised performances exploiting P2P policies to distribute and recover information. Figure 2 depicts the ab∗ This work was partially funded by the FET-IST Programme of the
European Commission, IST-2001-38113 MOBILE-MAN project.
Figure 1. The WB application interface
stractions we have used to support WB. The network level
provides basic connectivity among nodes through IP-like
routing and transport protocols. On top of them, a structured overlay network, comprising nodes that participate
in the WB application, is built. The overlay abstraction
is the fundamental substrate for any P2P application, providing functionalities such as logical node addressing (instead of topological, IP-like addressing) and subject-based
routing. Finally, an additional multicast level is used to efficiently distribute contents generated by application users
to all nodes in the overlay. These abstractions make quite
straightforward develop group communication applications.
They hide the complexity of low-level communications,
group management, and data distribution, and provide a robust, flexible, self-organising networking environment.
Figure 3 shows the complete networking solutions we
have used to support WB in real-world MANETs. We have
defined a first architecture (referred to as legacy), that uses
state-of-the-art components to implement the abstractions
in Figure 2. Specifically, it uses either AODV [1] or OLSR
[7] at the network level, Pastry [8] at the middleware level,
and Scribe [2] at the multicast level. While AODV and
OLSR represent standard models for ad hoc reactive and
proactive routing protocols, Pastry and Scribe have been designed for wired networks. The evaluation of the “legacy
solution” indicates weaknesses of these components, and
ways to improve them. In order to optimize the entire system performances, a cross-layer architecture, as depicted
on the right-hand side of Figure 3, has been proposed in
[4]. Specifically, the NeSt module allows cross-layer interactions between protocols at different layers. To this aim,
NeSt provides well-defined interfaces and data abstractions
to protocols [4], joining the advantages of cross-layering
Node A
Node D
Node C
Node B
node in the overlay
node NOT in the overlay
Figure 2. Abstractions supporting WB
Figure 4. Cross-layer interactions between
CrossROAD and OLSR
Scribe (XScribe)
Figure 3. Network solutions: legacy (left) and
cross layer (right)
and the scalability of traditional layered approach. CrossROAD represents an optimised solution at the middleware
layer that exploits cross-layer interactions with a proactive
routing protocol (OLSR in this case) in order to optimize
the creation and management of the overlay network. In
this paper we do not discuss any other MANET-optimised
solutions that could be integrated into the cross-layer architecture. However, in the framework of the MobileMAN
project, other such components both at the routing level
(Hazy Sighted Link State [9]), and at the multicast level
(X-layer Scribe) are being studied and currently under development.
3 Pastry vs. CrossROAD
Pastry represents the P2P computing model on which
CrossROAD has been designed to obtain great optimisations on ad hoc networks. It generates an overlay network by organising nodes in a circular logical address space
(ring). Specifically, it assigns to each node a logical identifier by hashing, for example, the node IP address. Logical identifiers determine the node position in the ring. In
addition, messages are routed over the ring by following
a subject-based policy, rather than a topology-based one.
Specifically, an application wishing to send a message m
has to provide a key k linked to m. The k value is hashed
to obtain an identifier in the same space of nodes’ logical
ids that is used to select the best destination for that message. The subject-based routing of Pastry sends the message to the node in the ring whose id is numerically closest
to the key hashed value. This policy represents the basis
for several distributed services; for example, Scribe exploits
subject-based routing to build and maintain multicast distribution trees.
To implement subject-based routing, Pastry builds at
each node a middleware routing table storing a subset of
other nodes’ ids. This table is initialised (during a bootstrap
phase) and updated (periodically) by exchanging information with the other nodes. When adopted in MANETs, this
approach generates quite a lot of network overhead. CrossROAD [5] provides the same Pastry functionalities through
the P2P commonAPI [3], but it drastically reduces the overlay management traffic by exploiting cross-layer interactions with a proactive routing protocol. Specifically, CrossROAD implements a Service Discovery protocol, that exploits the broadcast flooding of routing packets to distribute
services information. An example of cross-layer interaction between CrossROAD and OLSR is shown in Figure 4.
Each application running on CrossROAD has to register itself by specifying a service id (step 1). The list of service
ids registered at the local node (Node A in the figure) is
maintained by the Cross-Layer Plugin (XL-Plugin), which
can be seen as a portion of the NeSt module (step 2). The
XL-Plugin embeds the list of local service ids into periodic
Link-State Update packets generated by OLSR (step 3). On
the other nodes of the network (nodes B, C, D in the figure), upon receiving LSU packets containing such list, the
routing level notifies XL-Plugin to store the list in its internal data structures. This way, each CrossROAD node has
a complete knowledge of all the other nodes providing the
same service in the MANET, and it is able to autonomously
build the overlay network without generating any further
management traffic (step 5). Furthermore, in case of topology changes, the status of the overlay network converged as
quickly as the routing protocol does.
4 Experimental Results
The networking solutions described in Figure 3 have
been implemented and tested in a read-world multi-hop ad
hoc network. Specifically, the testbed consisted of 8 homogeneous laptops, out of which 6 run the WB application,
and the remaining 2 were used just as routers. Experiments
that have been run, which mimic the behavior of WB users
concurrently drawing strokes on their canvas. Users are represented by software agents that continuously interleave active phases (during which they draw a burst of strokes), and
idle phases (during which they just receive others’ bursts).
Idle phase durations and burst sizes are exponentially distributed. A traffic load of 100% is defined as the load generated by a user drawing – on average – 1 stroke per second.
Due to space constraints, we cannot provide here detailed measurements. Therefore, we discuss the outcomes
of some selected experiments, that allow us to highlight several benefits introduced by CrossROAD1 . Table 1 shows the
aggregate throughput (in the sending and receiving directions) of each node during the Pastry 80% and CrossROAD
100% experiments, respectively2. These results account for
the traffic generated from the routing up to the application
level. We mark node C as “C(R)” since it was the root of
the Scribe tree. Finally, the last two rows show the average
throughput computed over the nodes running WB including
and excluding C, respectively. Overall, when CrossROAD
is used instead of Pastry, the throughput is drastically reduced. The average value over all nodes in the CrossROAD
setup is about one third of the average value in the Pastry setup. It should be noted that, due to Scribe mechanisms, the root node has to handle a far greater amount
of application-level traffic than other nodes. Therefore, the
throughput reduction due to CrossROAD can be better emphasised by focusing on the last row of the table. If we
exclude node C, the average throughput in the CrossROAD
setup is about one fourth of the average throughput in the
Pastry setup. Finally, we found that CrossROAD also improves the stability of the Scribe tree. Table 2 shows the
number of sub-trees that are generated in Pastry and CrossROAD setup, respectively. When Pastry is used, the Scribe
tree is often partitioned in several isolated sub-trees, resulting in nodes to be isolated from the rest of the network.
Instead, this misbehavior is always avoided when CrossROAD is used. It can be shown that it is a byproduct of
the Pastry network overhead and bootstrap procedure.
5 Conclusions
In order to support P2P group-communication applications in MANETs, legacy network architectures designed
for wired networks are not the real solution. Specifically,
such solutions require too much management traffic, and
tend to saturate the scarce MANET resources. Optimis1 In the Pastry case, we herafter show only results from OLSR experiments, since OLSR generally allowed to achieve better performances than
2 We were not able to run Pastry experiments at 100% traffic load, because the testbed crashed due to excessive network load.
avg (no C)
Table 1. Throughput (Bps) in the Pastry 80%
and CrossROAD 100% setup.
80% (100%)
Table 2. Number of sub-trees at the Scribe
ing the network stack components through cross-layering
is a very promising way. In this paper, we have highlighted
drastic performance improvements by replacing Pastry with
CrossROAD, a cross-layer optimised P2P substrate. Further improvements might be envisaged if also the other P2P
components (e.g., Scribe) are optimised according to the
cross-layer paradigm.
[1] AODV, Dept. of Information technology at Uppsala University (Sweden), henrikl/aodv/.
[2] M. Castro, P. Druschel, A-M. Kermarrec and A. Rowstron,
“SCRIBE: A large-scale and decentralised application-level
multicast infrastructure”, IEEE Journal on Selected Areas in
Communication (JSAC), Vol. 20, No, 8, October 2002.
[3] F. Dabek and B. Zhao and P. Druschel and J. Kubiatowicz
and I. Stoica, “Towards a common API for Structured
Peer-to-Peer Overlays”, Proc. of the the 2nd International
Workshop on Peer-to-peer Systems (IPTPS’03), Berkeley,
CA, Feb. 2003.
[4] M. Conti, G. Maselli, G. Turi, “Design and evaluation of
a flexible cross-layer interface for ad hoc networks”, Proc.
of the Fourth Annual Mediterranean Ad Hoc Networking
Workshop (MedHocNet 2005), June 2005.
[5] F. Delmastro, “From Pastry to CrossROAD: Cross-layer
Ring Overlay for Ad hoc networks”, in Proc. of Workshop
of Mobile Peer-to-Peer 2005, in conjuction with the PerCom
2005 conference, Kauai Island, Hawaii, Mar. 2005.
[6] “Mobile Metropolitan Ad hoc Network (MobileMAN)”, IST2001-18113 Project, funded by the EC FET-IST Programme,
[7] OLSR, Andreas Tonnesen, Institute for informatics at the
University of Oslo (Norway),
[8] A. Rowstron and P. Druschel, “Pastry: Scalable, distributed
object location and routing for large-scale peer-to-peer
systems”, Middleware 2001, Germany, November 2001.
[9] C.A. Santivanez, I. Stavrakakis, and R. Ramanathan, ”Making
link-state routing scale for ad hoc networks”. In Proceedings
of the 2nd ACM Symposium on Mobile Ad Hoc Networking
and Computing (MOBIHOC’01), 2001.
Real Life Experience of Cooperation Enforcement Based on Reputation (CORE)
for MANETs
Claudio Lavecchia, Pietro Michiardi, Refik Molva1
Institut Eurecom, 2229, Route des Crêtes
06560 Valbonne Sophia Antipolis
{first name.last name}
Cooperation enforcement in mobile ad-hoc networks has
become a hot topic within the scientific community. Entities
belonging to a mobile ad-hoc network are prone to
selfishness because being cooperative and participating to
basic network functions such as routing and packet
forwarding involves resource consumption for the benefit of
others. Different approaches have been proposed to
promote cooperation in such environments. An
implementation of the cooperation enforcement mechanism
named CORE [5] is presented in the following as well as a
demonstration of its usage on a MANET testbed.
1. Introduction
The hype of trust establishment schemes that the research
community is witnessing in recent years results in a
proliferation of such mechanisms that target various issues
rising at different layers of a communication system.
A particular instance of trust establishment schemes is
represented by reputation mechanisms, in which the trust
metric takes the form of a reputation measure associated to
each entity taking part in a digital transaction. Reputation
can be defined as the level of trust inspired by entities based
on observations made on entities' past behavior. Intuitively,
reputation can be thought of as a metric that drives and
regulates the formation of dynamic communities that shares
interests and have common goals.
A typical setting in which reputation schemes are used to
regulate the formation and the survivability of digital
communities is represented by peer-to-peer (P2P) file
sharing systems. In P2P communities, reputation can be
used to baffle greediness and selfishness of peers that make
and use the system while at the same time suffer from the
dilemma of constrained resources. Indeed, is there a reason
to assume the volunteer participation to the community
welfare if no countermeasures are in place to stimulate a fair
distribution of the costs incurred by each individual to the
community operation? In general the answer is negative, as
it has been demonstrated by recent studies [2, 3]. Another
interesting domain of application of this type of trust
establishment schemes is offered by the mobile ad hoc
networking paradigm. In mobile ad hoc networks
(MANET), node participation to basic networking functions
such as routing and packet forwarding is of fundamental
importance. Recent studies [4] show that network
performance can be severely degraded even when only a
small fraction of the nodes that are part of an ad hoc
network deny participation to the network operation. Again,
scarce resources are at the origin of a selfish node behavior
whereby nodes (and end users operating those nodes) do not
want to share the (energetic) costs incurred by the network
operation for the benefit of others.
In this paper we focus on a reputation system used to
stimulate node participation to the execution of the packet
forwarding function in MANETs. We present an overview
of the CORE [5] reputation system architecture from an
implementation point of view and detail the demonstrative
setting in which we carried out the proof-of-concept
validation of CORE. The reader should refer to [6] for a
detailed description and analysis of CORE.
2. CORE System Architecture
The CORE reputation system uses the watchdog mechanism
[7]. A watchdog can be defined as a software component
installed on the nodes of a network with the aim of
observing neighboring nodes behavior with respect to
participation to basic network functions. The solution
proposed hereafter addresses only the packet forwarding
function and has been implemented and tested on a
MANET testbed made of nodes relying on off-the-shelf
802.11b hardware. A MANET node implementing the
watchdog mechanism must be able to overhear all the
packets that are sent within its wireless channel. To do so,
the 802.11b WLAN adapter needs to be operated in the socalled promiscuous mode. This functionality is
implemented at WLAN adapter firmware and driver level
and enables the WLAN adapter to pass all the packets
received to upper layers for further processing. In our
MANET testbed we use Dell TrueMobile 1150 WLAN
adapters that are operated by the Orinoco driver which
works in promiscuous mode out of the box. Once the
WLAN adapter is set in promiscuous mode, the packets are
captured using the pcap [8] C libraries. Those libraries are
available for Windows OS as well, making the porting
effort of the CORE mechanism to Windows OS acceptable.
The CORE reputation system has been implemented as a
Linux daemon, the implementation architecture is illustrated
This research was partially supported by the Information Society Technologies program of the European Commission, Future and Emerging
Technologies under the IST-2001-38113 MOBILEMAN project and by the Institut Eurecom.
in Figure 1. Here follows the detailed description of the
modules that compose the system:
Sniffer Module: monitors the packets that pass across layer
2 of the TCP/IP stack. This module passes the relevant
fields of packet headers to the analyzer module for further
analysis in the form of packet descriptors.
Analyzer Module: Receives packet descriptors from the
sniffer module and analyzes those descriptors to deduce
whether the neighbors are being cooperative or not.
The analyzer module includes an expectation table. Packet
descriptors that correspond to packets for which forwarding
is expected by a neighbor are stored in this table. The
scheduler included in the analyzer module triggers a
timeout each time that a packet descriptor is written in the
expectation table. Upon timeout expiration on a packet
descriptor, the analyzer verifies if the corresponding packet
has been forwarded from the neighbor to the next hop. In
such case it deduces that the neighbor that forwarded the
packet has been cooperative and a positive observation is
passed to the reputation module. If the packet has not been
forwarded before the timeout expiration, the node that was
expected to forward the packet is suspected to be selfish and
a negative observation is passed to the reputation module.
The analyzer module features an ARP interface that is
needed to perform some basic neighbor discovery functions.
Reputation Module: In the original version of CORE [5],
the reputation value associated to a node is evaluated in a
sophisticated way. The interested reader should refer to [6]
for a detailed description and analysis of advanced
reputation evaluation functions. For sake of simplicity and
in order to provide a proof of concept evaluation of CORE,
the real life implementation of our cooperation enforcement
mechanism is based on a simpler reputation function
described hereafter. The reputation module uses a weighted
average function to calculate reputation values for
neighbors according to observations provided by the
analyzer and stores those values in a reputation table. When
the reputation of a neighbor falls below a given threshold, it
issues punishment requests to the punishment module.
In the current implementation the reputation function is
given by:
k B1
R K a k 0
Obs a K k B
a is the node that is being observed by the watchdog.
B is the number of observation that the node that executes
the watchdog keeps in its local observation buffer.
W k is the weight given to the k-th observation.
K is the actual absolute discrete time.
Obs a K k is the value of the observation at time
K k
Possible observation values are: (+1) if the watchdog
detects a cooperative behavior, (-1) if the watchdog detects
a selfish behavior.
Punishment Module: Punishes selfish neighbors by
denying packet forwarding through the “iptables” Linux
framework [9].
The proposed architecture has the
advantage of identifying clear interfaces among the system
modules, thus allowing the interoperability of the watchdog
module with other systems that calculate and exploit the
reputation of other nodes, such as for example the one
proposed in [7]. Another advantage of this solution resides
in the very limited network overhead introduced by its
operation: the only additional traffic is a pair of ARP
request/reply generated each time that a node running the
watchdog analyzes packets that involve a previously
unknown neighbor node.
3. CORE Demonstration
The implementation of CORE has been integrated on our
MANET testbed. A demonstration has been developed to
show the system behavior.
The MANET testbed is composed of 4 nodes, equipped
with a WLAN adapter. Two of the nodes are laptops
running Windows OS, the third one, where the watchdog
software is executed is a laptop running Linux OS, the
fourth node can be either a laptop or a Compaq iPaq PDA,
in both cases it runs Linux OS.
The demonstration objectives are:
• To show how the CORE system maintains a reputation
state for all the neighbors of a MANET node. Two
states are possible for a neighbor: selfish and
• To show how the economical approach that drove the
CORE mechanism design effectively motivates a user
that is leaning toward selfishness to reconsider his
objectives and be cooperative.
• To show the effectiveness of inherent reintegration
mechanism proposed by CORE.
The MANET testbed nodes are logically disposed in a row.
We assume the existence of bidirectional wireless links
between neighbor nodes. Nodes are disposed as follows:
Ideally the physical distances between two non-neighboring
nodes of the network are larger that the WLAN card
transmission range. Two non-neighboring nodes of the
network that wish to communicate will pass through an
intermediate node that acts as a router. In reality, the
WLAN cards transmission ranges are so that the physical
separation in an indoor environment is achieved only when
two non-neighboring nodes are more than some tens of
meters far from each other and this is quite hard to handle in
a demo environment. For this reason we use some “tricks”
to logically separate two non-neighboring nodes. Those
tricks involve the usage of firewalls or Linux “iptables”
framework. Node W is the node that runs the watchdog.
This node observes the behavior of its neighbors. According
to the observations, node W maintains for each neighbor a
reputation state. State for a neighbor can be either
cooperative or selfish. With respect to the equation (1), the
watchdog parameters have been set as follows: B4 ,
W i 0 . 25 (all the weights are equal).
Node S is the node that is operated by the selfish user. As a
selfishness model, we assume that the administrator of this
node stops forwarding traffic originated by nodes A or W
and directed to node B and vice versa when the battery level
of the its device falls below a given threshold.
We assume a non-selective selfishness model that is
implemented by blocking the forwarding of other nodes
traffic through the usage of Linux “iptables” framework.
Node B serves the HTTP connection requests coming from
node W while node A acts as an FTP server for connection
requests coming from node S.
During the demonstration, node S changes its state from
cooperative to selfish. As a result of this change of attitude,
node W HTTP connection requests to node B will fail.
CORE console running on node W shows the observations
performed by the watchdog on the neighbors’ behavior.
When node S becomes selfish, a line appears on the CORE
console to show the neighbor change of attitude. As soon as
this transition is detected, the watchdog issues a punishment
request to the CORE punishment module. Node S is
immediately punished by node W, which stops forwarding
node S packets. From this moment on, the FTP connection
requests from node S to node A will stall.
When node S realizes that its FTP connection requests are
blocked by node W, it decides to become cooperative again.
This behavior transition is again captured by the watchdog
on node W and shown on the CORE console. Node W
immediately restarts forwarding node S packets. From this
moment node S is reintegrated in the network and its FTP
connections to node A will be successful again.
impact of different reputation functions on the system
[2] M. Feldman, C. Papadimitriou, J. Chuang, and I. Stoica, Free-Riding
and Whitewashing in Peer-to-Peer Systems, ACM SIGCOMM'04
Workshop on Practice and Theory of Incentives in Networked Systems
(PINS), August 2004
[3] Kevin Lai, Michal Feldman, Ion Stoica, John Chuang, Incentives for
Cooperation in Peer-to-Peer Networks, in Proceedings of Workshop on
Economics of Peer-to-peer Systems, June 5-6 2003
[4] Pietro Michiardi, Refik Molva, Simulation-based Analysis of Security
Exposures in Mobile Ad Hoc Networks, in Proceedings of European
Wireless 2002 Conference
[5] Pietro Michiardi, Refik Molva, CORE: A COllaborative REputation
mechanism to enforce node cooperation in Mobile Ad Hoc Networks, IFIP
CMS02, Communication and Multimedia Security Conference, September
[6] Pietro Michiardi, Cooperation enforcement and network security
mechanisms for mobile ad hoc networks, PhD Thesis
[7] Sergio Marti, T.J. Giuli, Kevin Lai, Mary Baker, Mitigating routing
misbehavior in mobile ad hoc networks, MobiCom00, International
Conference on Mobile Computing and Networking, 2000
[10] Sonja Buchegger, Jean-Yves Le Boudec, Performance Analysis of the
CONFIDANT Protocol: Cooperation Of Nodes Fairness In Dynamic Adhoc NeTworks, in Proceedings of IEEE/ACM Symposium on Mobile Ad
Hoc Networking and Computing (MobiHOC), 2002
ARP interface
4. Conclusion
To the best of our knowledge, this article presents the first
real-life implementation of a cooperation enforcement
mechanism based on reputation which does not have an
impact on underlying routing protocol adopted in the
MANET (as opposed to [10]). Cooperation enforcement
represents a fundamental building block for a heterogeneous
MANET where no a priori trust relationships can be
established among peers. The implementation of the
software component described in this work constitutes a
starting point for a thorough analysis (through
measurements) of the impact of a cooperation enforcement
mechanism on the operation of a heterogeneous MANET.
For our future work we plan to investigate the behavior of
the system in heavy load conditions to test the effectiveness
of promiscuous listening. We will work on the fine tuning
of the system parameters illustrated in (1) and study the
introduction of an adaptive observation sampling factor to
mitigate the CPU load generated by promiscuous listening
and improve the responsiveness of the system behavior to
the network conditions. We plan as well to investigate the
WLAN Driver
Figure 1. CORE Implementation Architecture
VoIP Testbed in Ad Hoc Networks
Jose Costa-Requena, Mohammand Ayyash, Jarrod Creado, Jarkko Hakkinen, Raimo Kantola and Nicklas Beijar
Networking Laboratory- Helsinki University of Technology
Otakaari 5, 02510 Espoo, P.O. Box 3000, Finland
Abstract— Ad hoc networks provide networking capabilites
without infrastructure support. Ad Hoc networks have been
researched for few years but still they do not get enough
interest form the industry. The reason is that they do not fulfil
any clear user needs. The actual wireless technologies already
provide users with enough coverage and the necessary network
capabilities for supporting high variety of peer to peer
applications. In this paper we present VoIP application on top of
Ad Hoc networks as a disruptive technology for voice
communications without infrastructure support. With Ad hoc
networks a low cost connectivity for VoIP, messaging and
presence services can be provided and we present the results
from a real VoIP testbed over Ad hoc networks.
Index terms—Ad Hoc, VoIP, Service Discovery
Voice communications have been a disruptive
technology in the past. Wireless telephony and Voice
over IP (VoIP) can be considered as the next disruptions
associated with voice communications. The next
telephony innovation is linked to new drivers such as
mobility and cost. In this paper we present the VoIP over
Ad hoc networks as the next disruption related to voice
communications. The feasibility of VoIP over Ad hoc
networks is demonstrated with real testbed results. The
results show the possibility of implementing voice
communications with a reasonable quality of service
(QoS). In order to implement the VoIP service we need
to glue together Ad hoc technology with the right
signalling and transport protocol.
network deployment. Ad hoc devices should perform
their own network topology functions keeping track of
the connection between nodes and performing routing
functionality. The link state in an Ad Hoc network
changes whenever the users move, and the nodes must
be able to provide automatic topology establishment and
maintenance. Ad Hoc nodes need to be self-contained
and have their own device discovery and control
paradigms. The nodes need a set of mechanisms to allow
a device to be automatically integrated and configured as
part of an Ad Hoc network.
In Ad Hoc networks a new paradigm based on ondemand routing, where the nodes search for the routes
only when needed, suits better than the existing fixed
routing mechanism, where the nodes maintain the
complete network topology. However, this routing does
not scale for large networks. Therefore, numerous
routing protocols and algorithms have been proposed.
However, the absence of performance data in non-trivial
network configurations continues to be a major problem.
The Ad hoc routing protocols are typically subdivided into
two main categories: proactive routing protocols and
reactive on-demand routing protocols.
Reactive on demand routing protocols such as AODV
[1], establish the route to a destination only when
demanded. Proactive routing protocols such as OLSR [2]
are derived from legacy Internet distance-vector and
link-state protocols. They attempt to maintain consistent
and updated routing information for every pair of
network nodes by propagating route updates.
The strength of an Ad Hoc network resides in the
growth of IP over wireless and the self-organized
networking feature that will enable pervasive and
ubiquitous computing. Ad Hoc networks are seen as a
suitable technology for embedded network devices in
multiple environments such as vehicles, sensors, mobile
telephones and personal appliances.
Ad Hoc networks present some new and unusual
challenges that had not been primary concerns in fixed
At the early days of mobile communications, operators
did not see relevant investing in mobile technologies.
Today VoIP is already the next disruption in voice
communications. Users are driven by cost and they start
using Voice over IP applications that provide cheap voice
communications. VoIP over Ad hoc networks will be the
next distruptive technology. However, the voice quality is
poor and users may not allow other people traffic to go
through their devices.
1st. VoIP architecture
VoIP services require a signalling protocol that supports
the routing and addressing. There are several protocols
used for initiating and controlling multimedia sessions and
in our testbed we have selected the Session Initiation
Protocol [3]. The media transport is exchanged using the
Real Time Transport protocol RTP [4].
A SIP session usually involves a User Agent (UA), a
Proxy Server, a Registrar Server, and a Location server.
Figure 1 represents a SIP registration flow diagram. The
session set up is initiated through the SIP Registrar that
can locate the destination SIP UA and acts as a proxy.
200 OK
Store User
User Lookup
180 Ringing
180 Ringing
The VoIP testbed we have developed uses
components from other research institutions [5]. In order
to increase the QoS in Ad hoc networks, enhancements
in the VoIP application have been implemented. The
experiment was carried out in the laboratories of the
Consiglio Nazionale delle Ricerche (IIT-CNR), Pisa,
Italy. The test logs and additional information can be
found in the project site at networking laboratory [6]. The
experiment measures the overall performance of audio
sessions using VoIP in Ad hoc wireless LAN
The VoIP client is runing in both Laptops and iPAQs
and additionally they require the following software: a
GSM library [7] and an RTP library [8]. The tests are
performed with two different routing protocols; OLSR
and ADOV. We analyse the routing protocol effect in
the overall performance.
200 OK
200 OK
Figure 1. SIP UA Registration and session setup.
Moreover, a SIP UA can obtain the destination IP
address using embedded service discovery mechanism
without network support. Figure 2 presents the flow
diagram for initiating a session in this case.
Jitter buffer length is adjusted to analyse the effect in
the quality of the audio session. Increasing it will reduce
the perceived pauses in audio playback, resulting in
smooth playback. On the other hand, this will increase
the overall delay. Different RTP payload is applied per
message, which is equivalent to increasing the amount of
audio data in each RTP packet. This will enhance the
audio playback at the receiver since each packet holds
enough audio data to play until the next packet arrives.
In addition, if one packet is lost, a larger mount of audio
data is lost resulting in longer pauses.
2nd. Test metrics
Table 1 shows the metrics used to analyse the
Embedded in
User Agent
Table 1. Performance test metrics
Obtain destination
User Agent IP addr
RTP Traffic
180 Ringing
200 OK
GSM data
Figure 2. Session setup without network support.
Overall QoS
Ratio between total RTP bytes sent divided by the
total bytes sniffed at the transmitter. This
represents the RTP traffic percentage of the total
channel capacity.
Ratio between total GSM bytes sent divided by the
total bytes sniffed by Ethereal tool. This represents
GSM traffic percentage of the total channel
capacity. GSM bytes are calculated by multiplying
the total number of RTP packets by the number of
GSM packets per RTP packet, and multiplied by the
GSM packet length, which is 33 bytes.
Ratio between total number of routing protocol sent
and received messages divided by the total number of
GSM bytes successfully sent.
Ratio of RTP packets successfully received out of
those intended to be transmitted.
those intended to be transmitted.
Loss in Link
Ratio of RTP packets lost in the link divided by the
RTP packets intended to be transmitted.
Jitter measured at the receiver during a continuous
reception time.
3rd. Test results
Table 2 presents results using OLSR and AODV
routing protocols with 3 GSM packets per RTP message,
where a 60ms and a 100ms GSM buffers are compared.
Table 2. OLSR and AODV results
4th.Test conclusions
The results show that when sending the same
percentage of RTP and GSM packets the signalling
overhead is bigger for OLSR (i.e. 3%). The OLSR has
to send periodically link state updates.
Increasing the GSM buffer in OLSR helps to
decrease the jitter. In case of a link broken AODV
reacts immediately and re-establishes the link. Thus, the
buffer helps to maintain a constant flow of audio packets.
However, in case of OLSR when the link is broken the
packet loss increases and the buffer has to store the
received packets until the link is re-establised.
T raffic
GSM data
Link Loss
Figure 3 and Figure 4 show the percentage of
signalling traffic versus voice data.
Legend: Non RTP Traffic, RTP Traffic, AODV Traffic
Figure 3. AODV signalling overhead
This paper presents a real VoIP testbed in Ad hoc
networks. The architecture proposes an integrated
routing functionality together with QoS optimisations for
voice sessions in the application layer. The work under
development consists of dynamically changing the
number of audio packets per RTP message. Therefore,
when the packet loss increases, the number of GSM
packets is reduced within each RTP message in order to
reduce the overall audio loss. Thus, when the link is
stable and the packet loss is lower, the number of GSM
packets increases in order to increment the audio packets
that reach the destination.
We would like to thank the MobileMAN members who
contributed to the testing in Pisa.
This work was partially funded by the Information
Society Technologies (IST) program of the European
Commission, Future and Emerging Technologies under
the IST 2001-38113 MOBILEMAN project.
This work has been partly supported by the European
Union under the E-Next Project FP6-506869
Legend: Non RTP Traffic , RTP Traffic, OLSR Traffic
Figure 4. OLSR signalling overhead
C. E. Perkins and E.M. Royer, "Ad-hoc On Demand Distance
Vector Routing", Second IEEE Workshop on Mobile Computing
Systems and Applications, pp. 90-100, February 1999.
P. Jacquet, P. Muhlethaler, A. Qayyum, A. Lanouiti, L. Viennot
and T. Clausen, "draft -ietf-MANET -olsr-02.txt", July 2000.
M. Handley, H. Schulzrinne, E Schrooler, J. Rosenberg, et. al.,.
"Session Initiation Protocol", RFC 3261, IET F.
H. Schulzrinne, S. Casner, R. Frederick and V. Jacobson, “RTP:
A Transport Protocol for Real-Time Applications,” RFC 1889.
AODV-UU, AODV implementation created at Uppsala
Universtity (
Helsinki University of Technology, Networking Laboratory
GSM library v06.10 (
RTP library v2.9 ( )
Experimenting a Layer 2-based Approach to Internet Connectivity for Ad Hoc
R. Bruno, M. Conti, E. Gregori, A. Pinizzotto
IIT Institute
National Research Council (CNR)
Via G. Moruzzi, 1 - 56124 PISA, Italy
Email: {r.bruno,m.conti,e.gregori,a.pinizzotto}
A prerequisite for the mass-market deployment of multihop ad hoc technologies is the capability of integrating with
existing wired infrastructure networks. However, current
solutions to support connectivity between ad hoc networks
and the Internet are based on complex mechanisms, such as
Mobile-IP and IP tunnelling. In this paper we propose a
lightweight solution based on simple Layer-2 mechanisms.
Experiments carried out in a real test-bed confirm the validity and efficiency of our approach.
1. Introduction
In spite of the massive efforts in researching and developing mobile ad hoc networks in the last decade, this type of
networks has not yet witnessed a widespread deployment.
The low commercial penetration of products based on ad
hoc networking technologies could be explained by considering that users are interested in general-purpose applications where high bandwidth and open access to the Internet
are consolidated and cheap commodities. For these reasons,
there is a growing interest in designing working mechanisms
for providing an easy access to the Internet to nodes in ad
hoc networks. In this paper, we address this issue proposing
a novel approach to ensure a working Internet connectivity to proactive ad hoc networks. We decided to consider
proactive routing protocols because this family of routing
protocols is more suitable than reactive protocols for supporting advanced services and applications in mobile ad hoc
Two classes of approaches have been proposed so far to
support connectivity between ad hoc networks and the Internet. One approach is to implement a Mobile IP Foreign
∗ This
work was funded by the Italian Ministry for Education and Scientific Research (MIUR) in the framework of the FIRB-VICOM project,
and by the Information Society Technologies program of the European
Commission under the IST-2001-38113 MobileMAN project.
E. Ancillotti
Dept. of Information Engineering
University of Pisa
Via Diotisalvi 2 - 56122 Pisa, Italy
Email: [email protected]
Agent (MIP-FA) in the ad hoc node that acts as Internet gateway, and to run Mobile IP in all the ad hoc nodes [1]. A different approach relies on the implementation of a Network
Address Translation (NAT) in the gateway [3], that translates the IP addresses of ad hoc nodes to an address on the
NAT gateway, which is routable on the external network.
Such approaches are based on complex IP-based mechanisms originally defined for the wired Internet, like the IP-inIP encapsulation, Mobile IP and explicit tunnelling, which
may introduce significant overheads. This paper proposes a
lightweight technique to provide global Internet connectivity to the ad hoc nodes, using only Layer-2 mechanisms.
The basic idea is to logically extend a wired LAN to the ad
hoc nodes in a transparent way for the wired nodes by developing a specific Proxy ARP daemon inside the gateway.
Experiments carried out in a real test-bed confirm the validity and efficiency of our approach.
2. Background on the OSLR Protocol
Our test-bed uses the OLSR protocol as ad hoc network
routing algorithm [2], although our solution can be applied
to any proactive scheme. The OLSR algorithm employs an
efficient dissemination of the network topology information
by selecting special nodes, the multipoint relays (MPRs),
to forward broadcast messages during the flooding process.
In order to allow the injection of external routing information into the ad hoc network, the OLSR protocol defines the
Host and Network Association (HNA) message. The HNA
message binds a set of network prefixes to the IP address
of the node attached to the external networks, i.e., the gateway node. In this way, each ad hoc node is informed of the
network address and netmask of a network that is reachable
through each gateway. Hence, OLSR exploits the mechanism of default routes to advertise Internet connectivity.
For instance, a gateway that advertises the default
route, will receive all the packets destined to IP addresses
without a route on the local ad hoc network.
3. Internet Connectivity using a Proxy ARP
The proposed architecture is depicted in Figure 1. An
OLSR-based ad hoc network is interconnected via a Master Gateway (MG) to a wired LAN, which provides the connectivity to the external Internet. The wired LAN is an IP
subnetwork identified by IP S /L, i.e., an IP network address (IP S ) and the network mask length L (for example
X.Y.96.0/22). The wired nodes’ IP addresses belong to
IP S /L. Thus, the wired hosts are able to exchange packets
using their ARP table, which is a list of mappings between
the IP address (Layer-3) and the MAC address (Layer-2) of
known hosts on the same subnet. The connectivity between
the wired LAN and the Internet is provided by a router and
standard IP routing protocols.
Figure 1. Test-bed implementation.
Our OLSR-based ad hoc network consist of: (i) mobile ad
hoc nodes and (ii) a fixed backbone formed by one MG node
and multiple Slave Gateways (SGs), connected via wired
links. The MG is a node with two wired interfaces and a
single wireless interface that acts as gateway allowing the
ad hoc network to be connected to the Internet. The SG
nodes have two interfaces, wireless and wired, which ensures the routing between separated parts of the ad hoc network through the wired infrastructure. The ad hoc routing protocol is implemented by an OLSRd daemon, which
runs in all the interfaces - wireless and wired - of mobile
nodes, SG nodes and MG node, with the exception of the
MG node’s wired interface IFext , which provides access
to the external network, advertising Internet connectivity as
default routes (i.e., through HNA messages). Mobile nodes
are configured with a static IP address belonging to the same
subnet of the wired LAN, i.e.,IP S /L. The hybrid nodes’
interfaces are configured with private IP addresses. This is
done to facilitate the configuration of the MG node, as explained later. It is worth remarking that we devised the SGs
nodes as simple entities supporting the ad hoc node mobility, and increasing the available bandwidth between the MG
node and the mobile nodes. For this reason they do not need
a globally routable IP address.
Connectivity for Outgoing Traffic. The OLSRd daemon builds the routing tables with entries that specify the IP
address of the next hop neighbour to contact to send a packet
destined to another host or subnetwork. Since the MG node
advertises as default route, all packets destined for
IP addresses without a route on the ad hoc network, will be
routed along the default route to the MG node and forwarded to the Internet. A special case is when a mobile node
wants to send a packet addressed to the node on the local
wired LAN (e.g., node H in Figure 1). As the destination IP
address, IPH , belongs to IP S /L, the routing table lookup
on the source node will assume that the destination node
is directly connected to the node wireless interface. This
will result in a failed ARP Request for the IPH address,
sent on the source node’s wireless interface. To solve this
problem, we split the original IP S /L subnet into two consecutive subnets, IP SL /(L+1) and IP SU /(L+1), such as
to have IP S /L = IP SL /(L+1) ∪ IP SU /(L+1). For example, X.Y.96.0/22 = X.Y.96.0/23∪X.Y.98.0/23. Thus,
the MG node has to advertise also these two subnets on HNA
messages. In this way, each mobile node will always have,
for any host H on the local wired LAN, a routing table entry
with a more specific network/mask than the one related to its
wireless interface (i.e., IP S /L). Consequently, the longestmatch criterion, in the routing table lookup, will determine
the right next hop for the IPH address.
Connectivity for Incoming Traffic. To allow the nodes
on the local wired LAN, including the router, to send packets to the mobile nodes, a specific daemon runs on the IFext
interface of the MG node, called Ad Hoc Proxy ARP daemon (AHPAd). This daemon periodically checks the Master
Gateway’s routing table and ARP table, such as to publish
the MG node’s MAC address for each IP address having an
entry in the routing table with a netmask
The entries related to the IFint interface of the MG node
and the SG nodes’ interfaces are excluded. Hence, the MG
node acts as a Proxy ARP for the mobile nodes. When an
host on the local wired LAN has to send a packet to a mobile node in the ad hoc network, it deems the destination on
its wired network. Then, the node checks its ARP table for
a IP-MAC mapping and, if it is not present it sends an ARP
Request. The MG node answers with an ARP reply providing its MAC address and the packets can be correctly sent to
the destination through the MG node.
3.1. Experimental Results
In our test-bed we used the OLSR UniK implementation for
Linux in version 0.4.8, adopting the default setting for each
protocol parameter. All nodes where located in the same
room, and the iptables feature of Linux was used to emulate
that all nodes were not in radio visibility of each other. To
1 hop
2 hop
3 hop
4 hop
Time (sec)
Figure 2. Throughput of a single TCP flow for
different chain length.
2 hop
3 hop
4. Concluding Remarks
To conclude this paper we discuss on limitations of the proposed approach and on the enhancements currently under
• On the MG node two wired interfaces are needed because the Proxy ARP does not allow to answer to ARP
Requests for IP addresses that are reachable through
the same interface on which the ARP Request was received. Nevertheless, they can be replaced by a single
wired interface and emulated by a bridging function
and two virtual interfaces.
and being within radio range of D (B), on 50 seconds time
interval (see Figure 1). To emulate a soft handoff, we also allowed the mobile node E (C) to have both C and D (A and B)
within radio range for 10 secons at the beginning of each interval. The experimental results are shown in Figure 3. The
curve with label “3 hops” refers to the node E’s mobility,
while the curve with label “2 hops” refers to the node C’s
mobility. The figure shows that the maximum throughput
the TCP connection achieves, is not affected by the mobility. This is due to the fact that our solution doesn’t need any
IP encapsulation to work, differently from [3]. However,
the TCP instability increases because the OLSR neighbour
sensing mechanism may take up to 6 seconds to discover the
link change.
Time (sec)
Figure 3. Throughput of a single TCP flow with
generate the asymptotic TCP traffic during the experiments
we used the iperf tool.
We performed a first set of experiments aimed at evaluating the impact on the TCP throughput of the number of
wireless hops traversed in the ad hoc network to reach the
MG node. The experimental results are shown in Figure 2.
As expected, the longer the route, the lower is the throughput achieved by the TCP flow, and the throughput decrease
follows an almost linear relationship. The figure shows also
that, although the nodes are static, the TCP throughput is
not stable, but the TCP flow could be in a stalled condition for several seconds. This can be explained considering
that the control frames are broadcast frames that are not acknowledged, hence more vulnerable to collisions and channel errors than unicast frames. As a consequence, losses of
control frames can induce the loss of valid routes.
The second set of experiments was carried out to verify
the impact of mobility on TCP throughput. We allowed node
E (C) to alternate between being within radio range of C (A)
• The proposed architecture is working only with proactive ad hoc routing protocols. In fact, the Proxy ARP
function on the MG node needs to know all the mobile
nodes’ IP addresses, which has to publish on its ARP
• Only one MG node, i.e., access point to Internet, is allowed in the ad hoc network to avoid conflicts between
different Proxy ARPs. The redirect mechanism could
be exploited to support multi-homed (i.e., with multiple MG nodes) ad hoc networks.
• The IP addressing inside the ad hoc network is static.
Address auto-configuration techniques should be integrated in our architecture.
[1] M. Benzaid, P. Minet, K. Al Agha, C. Adjih, and G. Allard. Integration of Mobile-IP and OLSR for a Universal Mobility. Wireless Networks, 10(4):377–388, July
[2] T. Clausen and P. Jaquet. Optimized Link State Routing
Protocol (OLSR). RFC 3626, October 2003.
[3] P. Engelstad, A. Tønnesen, A. Hafslund, and G. Egeland. Internet Connectivity for Multi-Homed Proactive
Ad Hoc Networks. In Proc. of ICC’2004, volume 7,
pages 4050–4056, Paris, France, June 20–24 2004.
Demo of Ana4: an Hybrid Local Area Ad hoc Network Architecture
Nicolas B OULICAULT, Guillaume C HELIUS and Eric F LEURY
INSA de Lyon – France
We present the implementation of Ana4, a practical
architecture suitable for interconnecting devices in an
as hoc hybrid network environment, where wired and
multi hop wireless technologies are used. Ana4 is a
2.5 ad hoc layer that allows a full compatibility with
To Internet
Figure 1. Ana4 test bed proposal.
Project description
In order to extend the coverage of actual wireless
devices, wireless ad hoc or mesh networks have been
proposed. The aim of this demo is to set up a small hybrid mesh network based on Ana4 [2, 3, 4, 5, 1], an interconnection architecture suitable for heterogeneous
and spontaneous networks that mix various kinds of
link layer technologies: wired and wireless multi hops.
By ad hoc architecture, we denote a set of rules and operations dealing with addressing and routing that must
be set up for the ad hoc network to offer basic services such as ad hoc connectivity, TCP/IP compatibility, Internet connectivity and vertical handoff support.
During this demo, we propose to set up an hybrid ad
hoc test bed based on Ana4, mixing wire and wireless
media and offering a full auto-configuration support, a
full TCP/IP compatibility and connection to the Internet
The Figure 1 illustrates the problem of interconnecting heterogeneous devices all together in an hybrid ad
hoc network. This topology will be used as the reference during the demo. Host A plays both the role of
gateway through its wire interface eth0 connected to
the Internet and the role of “base station” via its wireless interface wlan0 on a given channel 0. Host A
also runs a DHCP server and performs IP masquerad-
ing to connect the Internet to the ad hoc network. In
node A, only the wireless interface wlan0 is part of
the ad hoc network and configured in Ana4 ad hoc
mode. Host B has two wireless interfaces, one on
channel 0 and one on channel 1. Note that the two
link layer interfaces (wlan0 and wlan1) participate
to the ad hoc network and are gathered under the same
Ana4 virtual interface adh0. Hosts C and D both
have one wireless interface wlan0 and one Ethernet
interface eth0. For nodes C and D both wireless and
wire interfaces participate to the ad hoc network and
are gathered under one Ana4 interface adh0.
The Figure 2 illustrates the layers and abstractions
that we encounter in a generic Ana4 network. In our
test bed illustrated in Figure 1 we have at the physical
layer three separate physical connectivity graphs corresponding to channel 0 and channel 1 of the 802.11
link layer and to the Ethernet link. At the ad hoc layer
we merge this three graphs on an “ad hoc node” basis. That is, a node is no more a link layer 2 interface
but it is the gathering of all interfaces that participate
to the ad hoc network (all wireless interfaces/channels
and Ethernet cards in our case). At the IP layer, the ad
hoc network is seen as an Ethernet link.
The aim of this demo is to demonstrate the advantages of Ana4 and its compatibility with classical uses
abstracted ad hoc
IP level
Ad Hoc level
ad hoc network
802.11 channel 1
Hardware level
802.11 channel 0
Figure 2. The different abstractions in an
Ana4 network
of TCP/IP. For example, When host D joins the network, it uses a classical DHCP client to broadcast a
request on the local network in order to retrieve its IP
address, its DNS and its gateway. Despite the fact that
this local network is a multi- hop, multi-interface network, the broadcast packet is flooded inside the mesh
network thanks to Ana4 and the response is forwarded
hop by hop back to node D.
ad hoc network
from IP
to IP
Level 4
IP address
Level 3
ad hoc
ad hoc virtual
Level 2
ad hoc address
MAC address
from / to interfaces
Figure 3. The ad hoc virtual interface
The Ana4 architecture is based on the notion of vir-
tual ad hoc interface, illustrated in figure 3. The simple
but efficient principle behind the virtual interface is to
hide the different physical devices and hardware networks behind the illusion of a single virtual network.
At the ad hoc level, this virtual network is a wireless
multi-hop network; at the IP level, it is a switched Ethernet link. Another powerful characteristic of this architecture is to allow an host to use a device simultaneously in ad hoc and in classical modes. Suppose
that a physical device handled by a virtual ad hoc interface is also configured as an Internet device. From
the IP view, the mobile hosts two distinct interfaces.
IP networking is performed over these two interfaces
without interference. This will also be shown during
the demo.
[1] N. Boulicault, G. Chelius, and E. Fleury. Experiments
of ana4: An implementation of a 2.5 framework for
deploying real multi-hop ad hoc and mesh networks.
In IEEE ICPS Workshop on Multi-hop Ad Hoc Networks: from theory to reality (REALMAN 05), Santorini, Greece, July 2005.
[2] G. Chelius and E. Fleury. Ananas : A local area ad hoc
network architectural scheme. In MWCN 2002, Stockholm, Sweden, Sept 2002. IEEE.
[3] G. Chelius and E. Fleury. Ananas : A new adhoc network architectural scheme. RR 4354, INRIA, 2002.
[4] G. Chelius and E. Fleury. Design of an hybrid routing architecture. In Fifth IEEE International Conference on Mobile and Wireless Communications Networks (MWCN 2003), Singapor, October 2003. IEEE,
[5] G. Chelius and E. Fleury. Request for an ad hoc
addressing architecture. In Wireless Personal Multimedia Communications (WPNC), Abano Terme, Italy,
September 2004.
Haggle Architecture and Demo of its Real World Implementations
Pan Hui*, Jon Crowcroft*, James Scott#, Christophe Diot# ,Augustin Chaintreau# ,Richard Gass#
University of Cambridge
[email protected]
Intel Research Cambridge
{ james.w.scott, christophe.diot, augustin.chaintreau, richard.gass}
Pocket Switched Networks (PSN) are a new
communication model at the intersection of Mobile Adhoc Networks (MANET) and Delay Tolerant Networks
(DTN).We propose a general software architecture, the
Haggle Architecture for PSN, and demonstrate it using
a prototype including two PSN applications, namely
distributed file sharing and newsgroups.
1. Introduction
Mobile Ad-hoc Networks (MANET) research often
assumes network conditions including long-lived paths
and full connectivity, and is therefore targeted at
situations where the network is dense and static.
Previous work in Delay Tolerant Networks (DTN) [1]
has focused on partitioned networks with occasional
links between them, e.g. due to satellites or “message
ferries” [2]. Both models are not realistic when
considering our target application scenario, namely for
humans moving around during their daily lives. We
observe that, while humans often have end-to-end
connected paths via “islands” of Internet connectivity,
this is not always true, e.g. while traveling. It is also
unrealistic to assume that mobile humans will have a
wireless network path with other humans which they
wish to perform networking with. We proposed Pocket
Switched Networks (PSN) as a more realistic model for
human daily life. PSN is designed around real life
human mobility measurements [3][4] and using a topdown approach that is a practical and realistic
communication model based on what kind of
applications human need or may like to have in their
daily lives. PSN realizes the presences of gaps
between the “islands of connectivity” and makes use of
local connectivity and user mobility during these gaps,
and global connectivity for data delivery when it is
Haggle is a general architecture proposed for PSN
applications. It embraces “Application Layer Framing”
[5], with application-layer and network-layer
information collapsed into on space. It use Applicationlevel Data Units (ADU) as basic unit of communication
attribute-value tuples to
communication parties and different applications.
A distributed file sharing system and a newsgroup
making use of only human mobility and local
connectivity are implemented using the Haggle
Architecture. The implementation uses Java J9
platform running on PDA and PC and Bluetooth and
WiFi as underlying network connecting media. During
the demo, the Bluetooth version will be demonstrated.
2. Haggle Architecture
We now introduce Haggle, an implementation of
Pocket Switching which we are using to prototype
solutions to the challenges identified in Section 3. We
expect that the full specification and implementation of
Haggle will take several years. Development will
involve a process of iterative refinement, driven by
feedback from the deployment of trial applications. We
describe below the architectural elements that are
present in the initial prototype of Haggle.
2.1. Application-level Data Unit
In the spirit of “Application Layer Framing”
proposed by Clark and Tennenhouse in the early
nineties [5], the Haggle architecture is not layered;
application-layer and network-layer information are
collapsed into one space. We represent messages as
application-level data units (ADUs), which are
comprised of a number of attribute-value pairs.
A node decides whether to offer an ADUs to a
neighbour based on many attribute-value pairs, rather
than simply on the destination address as in IP-centric
networking. This is useful because, in Haggle, an
ADU’s destination can be loosely specified, or not
specified at all. A sender may not know the set of
devices available to the recipient. Alternatively, for
message types such as information queries, many
devices could take the role of destination.
A node might accept transmission of an ADU from a
neighbour for two reasons. One is that the node may
determine that it is a valid recipient, since it might
carry an application which is interested in the ADU.
The other is that the node might decide that it has a
high expectation of further forwarding opportunities for
that ADU, for example because the recipient was
recently seen by it, or because the node expects to
acquire global connectivity shortly, and can thereby
deliver the message. Again, the visibility of
application-level information in various attribute-value
pairs may usefully contribute to a node’s decision.
Examples of possible attribute-value pairs for various
Haggle applications are:
message-type: private message
originator: NAME James Scott
sender-authentication: [cryptographic signature]
recipient: NAME Pan Hui
recipient: SMTP [email protected]
recipient: SMS +441233764432
encrypted-data: [encrypted message]
message-type: query
data: opportunistic networking
message-type: public message
forum: recipes
subject: chicken madras
2.2. Node Architecture
The preliminary design of the Haggle node
architecture, illustrated in Figure 1, is based around the
notion of a Haggle Information Store (HIS) containing
ADUs on each user device. Applications can inject new
ADUs into the HIS, can register interest in incoming
ADUs by specifying matching criteria for particular
attribute-value pairs, and can search the HIS for
existing ADUs.
The Haggle control module decides whether to accept
an incoming ADU into the HIS, provides an API for
applications, and controls the forwarding module. The
forwarding module performs neighbour discovery
using the available network interfaces, and when
necessary connects to neighbours which are identified
as potential nexthops for ADUs in order to offer those
ADUs for transmission. Note that, where it is available,
the Internet is regarded as just one of many network
interface types.
Figure 1: Prototype Haggle Architecture
2.3.Implementing Haggle
We aim to develop Haggle for mobile computing
devices, which obviously includes notebook PCs and
PDAs. Recently, this term has also become applicable
to mobile phones, which now have significant storage,
computing power, local networking (generally in the
form of Bluetooth), and support for dynamicallyloaded applications.
We use Java in order to support a wide range of
platforms with a single reference codebase. For local
connectivity, we utilise Bluetooth and 802.11 in ad hoc
mode. Conveniently, Bluetooth is available on all our
targeted platforms. An initial step in the Haggle project
has been to develop implementations of the Java
Bluetooth standard (JSR-82 [6]) for Windows XP and
Windows CE, which we have released under the LGPL
open source license3.
3. Haggle Demo
The demos we given here are the ad-hoc file sharing
system and the ad-hoc newsgroup application. Both
applications are implemented based on the Haggle
architecture defined above. The implementation is done
on Java J9 platform on PDA with Bluetooth as the
transmission medium. Inside the demo, several PDAs
running the Haggle applications will be distributed to
the participants for testing.
3.1. File Sharing System
By using the Haggle File Sharing system, the users can
do searching and downloading files in a totally ad hoc
and opportunistic manner. Type a file name wanted to
search into the browser and a File Request ADU will
be sent to everyone encountered If a device contained
the file, it will generated a File Response ADU which
will be forward out in two manners 1) send to the
originator of the Request ADU when get into contact
range, or 2) send controlled copies into the devices
encountered which act as relays of this copy and would
send to the originator of the request when encountered.
4. Conclusions and Future Work
The Haggle Architecture for PSN is presented and
two haggle applications, file sharing system and
newsgroup are implemented on Java j9 platform
running on PDA with Bluetooth. We will make the
codes soon available under an open source license. The
implementation of WiFi connectivity is proceeding.
Over the next few years, we aim to build more
Haggle-based applications including instant messaging
and web browsing with automatic use of neighbours’
caches. We plan to test these applications by rolling out
prototypes to local users and deploying them to larger
user groups such as conference attendees. This will
enable us to study aspects of Haggle including usability,
scalability, network congestion, and user behaviour,
which can only be conclusively studied in deployments.
5. References
Figure 2: Haggle File Sharing User Interface
3.2. Newsgroup system
The Haggle Newsgroup allows the users to publish
to an ad-hoc newsgroup. When two devices see each
other, they will exchange ADU and the contents of the
newsgroup will be updated.
Figure 3: Haggle Newsgroup User Interface
[1] Kevin Fall. A delay-tolerant network architecture for
challenged internets. In Proceedings of ACM SIGCOMM,
[2] Wenrui Zhao, Mostafa Ammar, Ellen Zegura, "A
Message Ferrying Approach for Data Delivery in Sparse
Mobile Ad Hoc Networks," Proceedings of ACM Mobihoc
2004, Tokyo Japan, May 2004.
[3] A. Chaintreau, P. Hui, J. Crowcroft, C. Diot, R. Gass, and
J. Scott. Pocket switched networks: Real-world mobility and
its consequences for opportunistic forwarding. Technical
Report UCAM-CL-TR-617, University of Cambridge,
Computer Laboratory, February
[4] P. Hui, A. Chaintreau, J. Scott, R. Gass, J. Crowcroft, C.
Diot, Pocket Switched Networks and the Consequences of
t, To be presented at the Workshop on Delay Tolerant
Networking during SIGCOMM 2005.
[5] D. Clark and D. Tennenhouse. Architectural
considerations for a new generation of protocols. In ACM
SIGCOMM, pages 200–208, Sept. 1990.
Demo of residual bandwidth estimation in an 802.11 ad hoc network
Martin Nielsen
University of Oslo and UniK
Department of Informatics
Oslo, Norway
[email protected]
Quality of service in wireless LANs differ from their
wired counterpart in that variation in available resources
is not only due to competition by other traffic, but also due
to variation in link quality. Various proposals exist to remedy the effects of this dynamicity. Our aim is to implement
a system that by offering more information about the current per peer link state and resource usage can improve performance not only by provisioning through reservation, but
also in route calculation. To aid in achieving this, a closed
network has been constructed in order to conduct live experiments where outer factors like interference and other
competing stations are controlled.
1. Introduction
Though mobile ad hoc networks (MANETs) might arise
anywhere at any time one would not only want to support
reliable end-to-end communication, but also some kind of
premium service. For some scenarios of use, like battlefield deployment, it is a must. This complicates matters and
presents a need for some sort of management of available
and consumed resources. This is where quality of service
(QoS) comes to play.
To accommodate a premium service certain knowledge
about the surroundings has to be evident both in term of
available hosts and their capabilities. The effort on our part
is four fold:
• Sensing and probing in order to collect data concerning
the link state.
• Bandwidth estimation with metrics for transmission
and reception rates.
• Quality of service by adaptation of local flow policy
and providing better routing metrics. The ability of
the network to smooth things out and pick up general
throughput might be seen as a service improving the
general quality perceived by applications.
• Controlled live experiments. Though all the pieces
necessary for a live network are covered one would
still wish to control certain factors inherent to shared
mediums during development. These are the hidden
and the exposed node problem, node separation and
general interference from items like microwave ovens
and other appliances operating in the industrial, scientific and medical (ISM) bands that jam the communication in the vicinity of a receiver.
The demo will be of the two first objectives listed above;
an ad hoc network where bandwidth estimation is conducted in the driver where the passive estimation technique
implemented by CRC in Canada [4] is running alongside
our active probing implementation. This will be in combination with a walk through the inner workings of the implementation where we look at how hardware and driver
constraints impose limitations not observed in the safe and
predictable surroundings of NS-2 simulations.
The test bed will consist of three laptops with two of
them conversing and one listening and probing in order to
estimate residual channel capacity. In the following sections we will examine the components of the demo and the
implemented driver closer.
2. The link layer
It was argued in [3] that the 802.11 medium access control protocol (MAC) is not the best suited for ensuring QoS.
Because of the commercial proliferation of compliant hardware it has however become the de facto standard. Its usability can nevertheless be augmented by offering more and
better information to higher protocol layers. In that spirit
the following sections will discuss how to estimate residual
bandwidth and link quality in order to provide some useful
MAC metrics.
2.1. Estimating residual bandwidth
In contrary to wired networks, bandwidth estimation is
not a straightforward operation in shared medium radio networks. Though much research has been done and a theoretic
analysis of one hop bandwidth estimation is in place [2] it
is however not a simple task when multi hop scenarios and
real hardware have to be addressed.
There has always been a gap between the theoretical
analysis models and simulation results on one side and what
is possible with current commercial products on the other
side. Although results acquired by the former are indeed
useful, existing hardware and their drivers put constraints
on their applicability. This can be exemplified by looking
at how we can do bandwidth estimation on the offered devices. Two methods of doing exactly this, the active and the
passive estimation techniques, are described in the following sections. It must be noted however that these are neither
exclusive of each other nor the only proposals of how to
accomplish estimation.
2.2. Active probing
The active probing in [6] was designed mostly for one
hop networks with constant modulation during the estimation. Our effort has tried to augment that scheme to encompass multi hop scenarios and to allow for use of the modulation deemed best by the drivers own rate controller. To begin we will first describe how the original proposal worked
and then describe the additional considerations which had
to be done in order to accommodate our use.
In its infancy this method started out by sending out
a chain of probe packets to assess the channel occupancy. It essentially measures the time of each sequence
of RTS/CTS/DATA/ACK and attributes any delay in excess
of an estimated value to be caused by deferring from the
medium entailed by either physical or virtual carrier sensing. By uniformly distributing probes within every transmission interval one can equally represent the channel occupancy. Thereby it is possible to infer the channel state.
The authors suggested occupying 30 to 40 percent of the
channel with probes. The channel utilization is estimated
by applying the following formula:
Pi=Kc i
T − T p Kc
U = i=1 m
Where U is the utilization estimate, Tm is the measured MAC delay, T is the calculated average MAC delay when the channel is sensed idle, Kc is the number of
probes delayed more than predicted, K is the total number
of probes and ∆T is the interval between the probes. The
residual bandwidth is estimated by nres = (1 − U )nmax
where nmax is the theoretical maximum throughput that can
be calculated for a specific frame size, which implies that
residual bandwidth is dependent on frame size.
2.3. Adaptation of the active probing scheme
The active probing proposal does have some glitches.
The ones inherent when moving from a simulation environment to real hardware will be dealt with first, leaving the
comparative details when looking at it in conjunction with
the passive proposal described later.
Starting of with the universality of the method it has to
be mentioned that it has a strong coupling to the driver that
cannot be solved in user space when working under Linux.
The method must measure the time from passing the packet
onto the device till the exact time the ACK returns. To the
best of our knowledge only one current driver allows such
tight coupling, specifically the Atheros chipset MADwifi
driver [5]. However, this driver is not completely open since
its radio can be tuned to frequencies outside the ISM band,
the FCC and other similar authorities have pressed for keeping a part closed. This lowers accuracy but is still better than
the jiffy accuracy provided by the common wlan-ng-prism2
header found in most drivers.1
In addition the formula had to be custom tailored to the
multi rate probes that could be sent during measurement.
We do this by sorting the probing data into intervals based
on the rate used and calculating the mean value.
The limitations of the stock Linux kernel also put a constraint on accuracy. The standard timekeeping accuracy is
set to one millisecond for timed events. An addition is being developed to improve this [1], but has yet to be implemented by us. This gives us a slight problem because we
are not able to control the channel saturation to a tight specific level. This issue does not break the technique however
and is mostly a problem when trying to minimize saturation
problems. In addition we do not account for clock drift, but
as the time measured is per probe and not for the entire train
it should not be a source of error.
During the demo we will look closer at what actually
happens when sending a probe. Because we can see how
many retires, what kind of protection and the rate we can
reason closer about the events, and look if all the excess
delay can be attributed to cross traffic. This will illustrate
better the design issues for the simple probes. To se if the
algorithm responds to changes in traffic load two other machines will vary it so we can inspect the predicted residual
bandwidth. Though the conference room will undoubtedly
be full of other wireless signals we can still see if the values
are correlated.
1 The
information gathered by setting the interface in promiscuous or
rf-monitoring mode is time stamped, at least in the madwifi driver, with
jiffy precision. This is one ms in the 2.6 linux kernel which does not give
the precision needed.
2.4. Passive probing
In order to minimize the intrusion on the network one
can try to estimate residual bandwidth by listening to incoming traffic and adding to that channel occupancy seized
while sending. Because the driver does not grant access
to the network allocation vector (NAV) it is not possible to
directly estimate the residual bandwidth by simply monitoring the busy state given by the carrier sensing state machine.
What we can do instead is to take all the frames going in
and out of a node and the ones heard through promiscuous
mode to estimate how much time this has taken during an
interval. CRC in Canada has implemented this technique
in the Atheros chipset driver. The metrics provided by the
driver are per neighbour node to achieve greater granularity.
To be able to accurately depict the channel state one has
to pay great attention not only to the 802.11 frame format
but also to the sequence of frames. Additional consideration
has to be taken because one has to include the mean back
off times for RTS and data frames, and also non-standard
behaviour by the driver. We will try to send frames of different sizes and with various types of protection to see if we
notice the difference. This is also a useful technique to see
if the driver complies with our requests as other types of experiments can get interesting results when based on flawed
2.5. Complementary performance
The two techniques both have their limitations. The active probing saturates the channel from 30 to 40 percent and
is not good when load is heavy. It could also be a problem if
two nodes within range probe at the same time. But in contrast to the passive technique it can estimate the intersection
of the bandwidth at the sender and the peer. The peer can be
overloaded by traffic we cannot hear and probing is a way
to estimate this. As mentioned we do not have access to the
NAV so it is not possible to estimate how often the physical carrier sensing is busy. Active probing defers from the
channel and takes this into account.
We propose combining the two methods to harvest their
benefits and reduce their negative sides. If the channel saturation is sensed above 50 percent one could argue that probing would be unwise and if no good information is available
the active method is the way to go.
Since the nodes are to close to each other during the
demo we will not be able to simulate a node chain showing how the two methods can complement each other. But
we can run them in parallel and compare their results. This
could perhaps shed some more light on what they encompass and where they lack.
3. Conclusion
This is still very much a work in progress where minimizing error is a major concern. As this is simply a service layer it is important to create a valuable interface with
meaningful metrics so that traffic shapers, routing protocols
and admission controllers can achieve a significant boost.
Otherwise the effort is only of theoretical interest. In addition we are considering augmenting the interface with received signal strength indicator and expected transmission
count to give a better picture of link quality.
4. Acknowledgements
I would like to thank Lars Landmark for providing comments on my work and the demo proposal, and Frank Li for
his insights to the active probing technique. Also I would
like to especially thank Mathieu Deziel for being so forthcoming with regards to the passive estimation technique.
[1] G. Anzinger.
High resolution POSIX timers.
[2] G. Bianchi. Performance analysis of the IEEE 802.11 distributed coordination function. IEEE journal on selected areas in communications, 18(3), 2000.
[3] L. Chen and W. B. Heinzelman. Qos-aware routing based
on bandwidth estimation for mobile ad hoc networks. IEEE
journal on selected areas in communications, 23(3), March
[4] M. Deziel and L. Lamont. Implementation of an IEEE 802.11
link available bandwidth algorithm to allow cross-layering.
[5] S. Leffler, M. Renzmann, and G. Chesson. Multiband atheros
driver for wifi.
[6] F. Y. Li, O. Kure, P. Spilling, M. Hauge, and A. Hafslund.
Estimating residual bandwidth in 802.11-based ad hoc networks: An empirical approach, September 2004.

Similar documents


Report this document