PSR: A Lightweight Proactive Source Routing Protocol For Mobile

Document technical information

Format pdf
Size 1.3 MB
First found Nov 13, 2015

Document content analysis

Language
English
Type
not defined
Concepts
no text concepts found

Persons

Emily D. West
Emily D. West

wikipedia, lookup

Morris R. Jeppson
Morris R. Jeppson

wikipedia, lookup

Henry St John, 1st Viscount St John
Henry St John, 1st Viscount St John

wikipedia, lookup

Cheng Li-wen
Cheng Li-wen

wikipedia, lookup

Organizations

Places

Transcript

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 63, NO. 2, FEBRUARY 2014
859
PSR: A Lightweight Proactive Source Routing
Protocol For Mobile Ad Hoc Networks
Zehua Wang, Student Member, IEEE, Yuanzhu Chen, Member, IEEE, and Cheng Li, Senior Member, IEEE
Abstract—Opportunistic data forwarding has drawn much
attention in the research community of multihop wireless networking, with most research conducted for stationary wireless
networks. One of the reasons why opportunistic data forwarding
has not been widely utilized in mobile ad hoc networks (MANETs)
is the lack of an efficient lightweight proactive routing scheme
with strong source routing capability. In this paper, we propose a lightweight proactive source routing (PSR) protocol. PSR
can maintain more network topology information than distance
vector (DV) routing to facilitate source routing, although it has
much smaller overhead than traditional DV-based protocols [e.g.,
destination-sequenced DV (DSDV)], link state (LS)-based routing
[e.g., optimized link state routing (OLSR)], and reactive source
routing [e.g., dynamic source routing (DSR)]. Our tests using computer simulation in Network Simulator 2 (ns-2) indicate that the
overhead in PSR is only a fraction of the overhead of these baseline
protocols, and PSR yields similar or better data transportation
performance than these baseline protocols.
Index Terms—Differential update, mobile ad hoc networks
(MANETs), opportunistic data forwarding, proactive routing,
routing overhead control, source routing, tree-based routing.
I. I NTRODUCTION
A
mobile ad hoc network (MANET) is a wireless communication network, where nodes that are not within the
direct transmission range of each other require other nodes to
forward data. It can operate without existing infrastructure and
support mobile users, and it falls under the general scope of
multihop wireless networking. This networking paradigm originated from the needs in battlefield communications, emergency
operations, search and rescue, and disaster relief operations. It
has more recently been used for civilian applications such as
community networks. A great deal of research results have been
published since its early days in the 1980s [1]. The most salient
research challenges in this area include end-to-end data transfer,
link access control, security, and providing support for real-time
multimedia streaming [2].
Manuscript received January 11, 2013; revised May 6, 2013; accepted June 22,
2013. Date of publication August 21, 2013; date of current version February 12,
2014. This work was supported in part by the Natural Sciences and Engineering
Research Council of Canada under Discovery Grant 293264-12, Discovery
Grant 327667-2010, and Strategic Project Grant STPGP 397491-10. The review
of this paper was coordinated by Prof. N. Kato.
Z. Wang is with Electrical and Computer Engineering, University of British
Columbia, Vancouver BC V6T 1Z4, Canada (e-mail: [email protected]).
Y. Chen is with the Department of Computer Science, Memorial University
of Newfoundland, St. John’s, NL A1B 3X5, Canada (e-mail: [email protected]).
C. Li is with the Faculty of Engineering and Applied Science, Memorial University of Newfoundland, St. John’s, NL A1B 3X5, Canada (e-mail:
[email protected]).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TVT.2013.2279111
The network layer has received a great deal of attention in the
research on MANETs. As a result, abundant routing protocols
in this network with differing objectives and for various specific
needs have been proposed [3]. In fact, the two most important operations at the network layer, i.e., data forwarding and
routing, are distinct concepts. Data forwarding regulates how
packets are taken from one link and put on another. Routing
determines what path a data packet should follow from the
source node to the destination. The latter essentially provides
the former with control input. Despite the amount of effort in
routing in ad hoc networks, data forwarding, in contrast, follows
the same paradigm as in Internet Protocol (IP) forwarding in the
Internet. IP forwarding was originally designed for multihop
wired networks, where one packet transmission can be only
received by nodes attached to the same cable. However, in
wireless networks, when a packet is transmitted over a physical
channel, it can be that channel. Traditionally, overhearing a
packet not intended for the receiving node had been considered
completely negative, i.e., interference. Thus, in a sense, the goal
of the research in wireless networking was to make wireless
links as good as wired links.
Opportunistic data forwarding represents a promising solution to utilize the broadcast nature of wireless communication
links [4]. Opportunistic data forwarding refers to a way in
which data packets are handled in a multihop wireless network. Unlike traditional IP forwarding, where an intermediate node looks up a forwarding table for a dedicated next
hop, opportunistic data forwarding allows potentially multiple
downstream nodes to act on the broadcast data packet. One of
the initial works on opportunistic data forwarding is selective
diversity forwarding by Larsson [5]. In this paper, a transmitter
picks the best forwarder from multiple receivers, which successfully received its data, and explicitly requests the selected
node to forward the data. However, its overhead needs to be
significantly reduced before it can be implemented in practical
networks. This issue was successfully addressed in the seminal
work on ExOR [6], outlining a solution at the link and network
layers. In ExOR, nodes are enabled to overhear all packets on
the air; therefore, a multitude of nodes can potentially forward
a packet as long as they are included in the forwarder list
carried by the packet. By utilizing the contention feature of the
medium-access-control (MAC) sublayer, the forwarder closer
to the destination will access the medium more aggressively.
Therefore, the MAC sublayer can determine the actual next-hop
forwarder to better utilize the long-haul transmissions.
To support opportunistic data forwarding in a mobile wireless network as in ExOR, an IP packet needs to be enhanced
such that it lists the addresses of the nodes that lead to the
0018-9545 © 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
860
IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 63, NO. 2, FEBRUARY 2014
packet’s destination. This entails a routing protocol where
nodes see beyond merely the next hop leading to the destination. Therefore, link state (LS) routing [e.g., optimized LS
routing (OLSR)] or source routing [e.g., dynamic source routing (DSR)] would seem to be good candidates. On one hand,
LS routing protocols include interconnectivity information between remote nodes, which is hardly useful for a particular
source node, but this incurs prohibitively large overhead. This
is even true with optimization techniques such as multipoint
relaying, as in OLSR [7]. On the other hand, if we wish to support opportunistic data forwarding in a MANET with constantly
active data communication between many node pairs, the reactive nature of DSR [8] renders it unsuitable. Meanwhile, source
routing is able to tightly control data forwarding paths. Thus,
it is not only of interest in opportunistic data forwarding but
also in a wider scope such as avoiding congestion, bypassing
malicious nodes, and allocating network resources.
In this paper, we propose a lightweight proactive source routing (PSR) protocol to facilitate opportunistic data forwarding in
MANETs. In PSR, each node maintains a breadth-first search
spanning tree of the network rooted at itself. This information is
periodically exchanged among neighboring nodes for updated
network topology information. Thus, PSR allows a node to
have full-path information to all other nodes in the network,
although the communication cost is only linear to the number
of the nodes. This allows it to support both source routing and
conventional IP forwarding. When doing this, we try to reduce
the routing overhead of PSR as much as we can. Our simulation
results indicate that PSR has only a fraction of overhead of
OLSR, DSDV, and DSR but still offers a similar or better data
transportation capability compared with these protocols.
The remainder of this paper is organized as follows. Section II reviews related work on routing protocol in MANETs.
Section III describes the design and implementation details
of our proposed routing scheme. The computer simulation,
related experiment settings, and comparisons between PSR and
existing algorithms are presented in Section IV. Section V
concludes this paper with a discussion of future research.
II. R ELATED W ORK
Routing protocols in MANETs can be categorized using an
array of criteria. The most fundamental among these is the timing of routing information exchange. On one hand, a protocol
may require that nodes in the network should maintain valid
routes to all destinations at all times. In this case, the protocol
is considered proactive, which is also known as table driven.
Examples of proactive routing protocols include destinationsequenced distance vector (DSDV) [9] and OLSR [7]. On the
other hand, if nodes in the network do not always maintain
routing information, when a node receives data from the upper
layer for a given destination, it must first find out about how to
reach the destination. This approach is called reactive, which is
also known as on demand. DSR [8] and ad hoc on-demand DV
(AODV) [10] fall in this category.
These well-known routing schemes can be also categorized
by their fundamental algorithms. The most important algorithms in routing protocols are DV and LS algorithms. In an LS,
every node floods the information of the links between itself and
its neighbors in the entire network, such that every other node
can reconstruct the complete topology of the network locally. In
a DV, a node only provides its neighbors with the cost to reach
each destination. With the estimates coming from neighbors,
each node is able to determine which neighbor offers the best
route to a given destination. Both LS and DV support the vanilla
IP packets. DSR, however, takes a different approach to ondemand source routing. In DSR, a node employs a path search
procedure when there is a need to send data to a particular
destination. Once a path is identified by the returning search
control packets, this entire path is embedded in each data
packet to that destination. Thus, intermediate nodes do not
even need a forwarding table to transfer these packets. Because
of its reactive nature, it is more appropriate for occasional or
lightweight data transportation in MANETs.
AODV, DSDV, and other DV-based routing algorithms were
not designed for source routing; hence, they are not suitable
for opportunistic data forwarding. The reason is that every
node in these protocols only knows the next hop to reach a
given destination node but not the complete path. OLSR and
other LS-based routing protocols could support source routing,
but their overhead is still fairly high for the load-sensitive
MANETs. DSR and its derivations have a long bootstrap delay
and are therefore not efficacious for frequent data exchange,
particularly when there are a large number of data sources.
In fact, many lightweight routing protocols had been proposed for the Internet to address its scalability issue, i.e., all
naturally “table driven.” The path-finding algorithm (PFA) [11]
is based on DVs and improves them by incorporating the
predecessor of a destination in a routing update. Hence, the
entire path to each node can be reconstructed by connecting
the predecessors and destinations; therefore, the source node
will have a tree topology rooted at itself. In the meantime, the
link vector (LV) algorithm [12] reduces the overhead of LS
algorithms to a great deal by only including links to be used
in data forwarding in routing updates. The extreme case of LV,
where only one link is included per destination, coincides with
the PFA.
PFA and LV were both originally proposed for the Internet,
but their ideas were later used to devise routing protocols in
the MANET. The Wireless Routing Protocol (WRP) [13] was
an early attempt to port the routing capabilities of LS routing
protocols to MANETs. It is built on the same framework of the
PFA for each node to use a tree to achieve loop-free routing.
Although it is an innovative exploration in the research on
MANETs, it has a rather high communication overhead due
to the amount of information stored at and exchanged by the
nodes. This is exacerbated by the same route update strategy
as in the PFA, where routing updates are triggered by topology
changes. Although this routing update strategy is reasonable for
the PFA in the Internet, where the topology is relatively stable,
this turns out to be fairly resource demanding in MANETs. (Our
original intention was to include the WRP in the experimental
comparison later in this paper, and we have implemented WRP
in ns2. Unfortunately, our preliminary tests indicate that the
changing topology in the MANET incurs an overwhelming
amount of overhead, i.e., at least an order of magnitude higher
WANG et al.: PSR
than the other mainstream protocols. Thus, we do not include
the simulation result of WRP as a baseline in our comparison.)
The PSR protocol proposed in this paper uses tree-based
routing as in PFA and WRP. To make our PSR more suitable
for the MANETs, we adopt a combined route update strategy
that takes advantage of both “event-driven” and “timer-driven”
approaches. Specifically, nodes would hold their broadcast after
receiving a route update for a period of time. If more updates
have been received in this window, all updates are consolidated
before triggering one broadcast. The period of the update cycle
is an important parameter in PSR. Furthermore, we go an extra
mile to reduce its routing overhead. First, we interleave fulldump and differential updates to strike the balance between
efficient and robust network operation. Second, we package
affected links into forests to avoid duplicating nodes in the
data structure. Finally, to further reduce the size of differential
update messages, each node tries to minimize the alteration of
the routing tree that it maintains as the network changes its
structure.
III. D ESIGN OF P ROACTIVE S OURCE ROUTING
Essentially, PSR provides every node with a breadth-first
spanning tree (BFST) of the entire network rooted at itself.
To do that, nodes periodically broadcast the tree structure to
their best knowledge in each iteration. Based on the information
collected from neighbors during the most recent iteration, a
node can expand and refresh its knowledge about the network
topology by constructing a deeper and more recent BFST. This
knowledge will be distributed to its neighbors in the next round
of operation (see Section III-A). On the other hand, when a
neighbor is deemed lost, a procedure is triggered to remove
its relevant information from the topology repository maintained by the detecting node (see Section III-B). Intuitively,
PSR has about the same communication overhead as DV-based
protocols. We go an extra mile to reduce the communication
overhead incurred by PSR’s routing agents. Details about this
overhead reduction will be discussed in Section III-C.
Before describing the details of PSR, we will first review
some graph-theoretic terms used here. Let us model the network
as undirected graph G = (V, E), where V is the set of nodes
(or vertices) in the network, and E is the set of wireless links (or
edges). Two nodes u and v are connected by edge e = (u, v) ∈
E if they are close to each other and can directly communicate
with given reliability. Given node v, we use N (v) to denote
its open neighborhood, i.e., {u ∈ V |(u, v) ∈ E}. Similarly, we
use N [v] to denote its closed neighborhood, i.e., N (v) ∪ {v}
(see [14] for other graph-theoretic notions).
A. Route Update
Due to its proactive nature, the update operation of PSR is
iterative and distributed among all nodes in the network. At
the beginning, node v is only aware of the existence of itself;
therefore, there is only a single node in its BFST, which is root
node v. By exchanging the BFSTs with the neighbors, it is able
to construct a BFST within N [v], i.e., the star graph centered at
v, which is denoted Sv .
861
In each subsequent iteration, nodes exchange their spanning
trees with their neighbors. From the perspective of node v,
toward the end of each operation interval, it has received a set
of routing messages from its neighbors packaging the BFSTs.
Note that, in fact, more nodes may be situated within the
transmission range of v, but their periodic updates were not
received by v due to, for example, bad channel conditions.
After all, the definition of a neighbor in MANETs is a fickle
one. (We have more details on how we handle lost neighbors
subsequently.) Node v incorporates the most recent information
from each neighbor to update its own BFST. It then broadcasts
this tree to its neighbors at the end of the period. Formally, v
has received the BFSTs from some of its neighbors. Including
those from whom v has received updates in recent previous
iterations, node v has a BFST, which is denoted Tu , cached for
each neighbor u ∈ N (v). Node v constructs a union graph, i.e.,
Gv = Sv ∪
(Tu − v).
(1)
u∈N (v)
Here, we use T − x to denote the operation of removing the
subtree of T rooted at node x. As special cases, T − x = T if
x is not in T , and T − x = ∅ if x is the root of T . Then, node
v calculates a BFST of Gv , which is denoted Tv , and places Tv
in a routing packet to broadcast to its neighbors.
In fact, in our implementation, the given update of the BFST
happens multiple times within a single update interval so that
a node can incorporate new route information to its knowledge
base more quickly. To the extreme, Tv is modified every time
a new tree is received from a neighbor. Apparently, there is
a tradeoff between the routing agent’s adaptivity to network
changes and computational cost. Here, we choose routing adaptivity as a higher priority assuming that the nodes are becoming
increasingly powerful in packet processing. Nevertheless, this
does not increase the communication overhead at all because
one routing message is always sent per update interval.
Assume that the network diameter, i.e., the maximum pairwise distance, is D hops. After D iterations of operation,
each node in the network has constructed a BFST of the
entire network rooted at itself since nodes are timer driven
and, thus, synchronized. This information can be used for any
source routing protocol. The amount of information that each
node broadcasts in an iteration is bounded by O(|V |), and the
algorithm converges in D iterations.
B. Neighborhood Trimming
The periodically broadcast routing messages in PSR also
double as “hello” messages for a node to identify which other
nodes are its neighbors. When a neighbor is deemed lost, its
contribution to the network connectivity should be removed;
this process is called neighbor trimming. Consider node v. The
neighbor trimming procedure is triggered at v about neighbor u
either by the following cases:
1) No routing update or data packet has been received from
this neighbor for a given period of time.
2) A data transmission to node u has failed, as reported by
the link layer.
862
IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 63, NO. 2, FEBRUARY 2014
Fig. 1. Binary tree.
Node v responds by:
1) first, updating N (v) with N (v) − {u};
2) then, constructing the union graph with the information
of u removed, i.e.,
Gv = Sv ∪
(Tw − v)
(2)
w∈N (v)
3) finally, computing BFST Tv .
Notice that Tv , which is thus calculated, is not broadcast
immediately to avoid excessive messaging. With this updated
BFST at v, it is able to avoid sending data packets via lost
neighbors. Thus, multiple neighbor trimming procedures may
be triggered within one period.
C. Streamlined Differential Update
In addition to dubbing route updates as hello messages in
PSR, we interleave the “full dump” routing messages, as stated
previously, with “differential updates.” The basic idea is to send
the full update messages less frequently than shorter messages
containing the difference between the current and previous
knowledge of a node’s routing module. Both the benefit of this
approach and balancing between these two types of messages
have been extensively studied in earlier proactive routing protocols. In this paper, we further streamline the routing update
in two new avenues. First, we use a compact tree representation
in full-dump and differential update messages to halve the size
of these messages. Second, every node attempts to maintain an
updated BFST as the network changes so that the differential
update messages are even shorter.
1) Compact tree representation. For the full-dump messages,
our goal is to broadcast the BFST information stored at
a node to its neighbors in a short packet. To do that,
we first convert the general rooted tree into a binary
tree of the same size, e.g., s nodes, using left-child
sibling representation. Then, we serialize the binary tree
using a bit sequence of 34 × s bits, assuming that IPv4
is used. Specifically, we scan the binary tree layer by
layer. When processing a node, we first include its IP
address in the sequence. In addition, we append two
more bits to indicate if it has the left and/or right child.
For example, the binary tree in Fig. 1 is represented as
A10B11C11D10E00F00G11H00I00. As such, the size of
the update message is a bit over half compared with
the traditional approach, where the message contains a
discrete set of edges.
The difference between two BFSTs can be represented
by the set of nodes that have changed parents, which are
essentially a set of edges connecting to the new parents.
We observe that these edges are often clustered in groups.
That is, many of them form a sizeable tree subgraph of
the network. Similar to the case of full dump, rather than
using a set of loose edges, we use a tree to package the
edges connected to each other. As a result, a differential
update message usually contains a few small trees, and its
size is noticeably shorter.
2) Stable BFST. The size of a differential update is determined by how many edges it includes. Since there can
be a large number of BFSTs rooted at a given node of
the same graph, we need to alter the BFST maintained
by a node as little as possible when changes are detected.
To do that, we modify the computation described earlier
here, such that a small portion of the tree needs to change
either when a neighbor is lost or when it reports a new
tree.
Consider node v and its BFST Tv . When it receives
an update from neighbor u, which is denoted Tu , it
first removes the subtree of Tv rooted at u. Then, it
incorporates the edges of Tu for a new BFST. Note that
the BFST of (Tv − u) ∪ Tu may not contain all necessary
edges for v to reach every other node. Therefore, we still
need to construct union graph
(Tv − u) ∪
(Tw − v)
(3)
w∈N (v)
before calculating its BFST. To minimize the alteration
to the tree, we add one edge of Tw − v to Tv − u at a
time. When node v thinks that a neighbor u is lost, it
deletes edge (u, v) but still utilizes the network structure
information contributed by u earlier. That is, even if it has
moved away from v, node u may still be within the range
of one of v’s neighbors. As such, Tv should be updated to
a BFST of
(Tv − u) ∪ (Tu − v) ∪
(Tw − v).
(4)
w∈N (v)
Note that, since N (v) no longer contains u, we need
to explicitly put it back into the equation. Similarly in
this case, the edges of (Tu − v) ∪ w∈N (v) (Tw − v) are
added to (Tv − u) one at a time, with those just removed
because of u taking priority.
IV. P ERFORMANCE E VALUATION
We study the performance of PSR using computer simulation
with Network Simulator 2 version 2.34 (ns-2). We compare
PSR against OLSR [7], DSDV [9], and DSR [8], which are
three fundamentally different routing protocols in MANETs,
with varying network densities and node mobility rates. We
measure the data transportation capacity of these protocols
supporting the Transmission Control Protocol (TCP) and the
WANG et al.: PSR
863
User Datagram Protocol (UDP) with different data flow deployment characteristics. Our tests show that the overhead of
PSR is indeed only a fraction of that of the baseline protocols.
Nevertheless, as it provides global routing information at such
a small cost, PSR offers similar or even better data delivery
performance. Here, we first describe how the experiment scenarios are configured and what measurements are collected.
Then, we present and interpret the data collected from networks
with heavy TCP flows and from those with light UDP streams.
A. Experiment Settings
Since many routing protocols’ performances are well known
in the classic two-ray ground reflection propagation model, we
select such a model as well in our simulation to present a
consistent and comparable result.1 Without loss of generality,
we select a 1-Mb/s nominal data rate at the IEEE 802.11 links
to study the relative performance among the selected protocols.
With the default physical-layer parameters of the simulator,
the transmission range is approximately 250 m, and the carrier
sensing range is about 550 m.
We compare the performance of PSR with that of OLSR,
DSDV, and DSR. The reasons that we select these baseline
protocols that are different in nature are as follows. On one
hand, OLSR and DSDV are both proactive routing protocols,
and PSR is also in this category. On the other hand, OLSR
makes complete topological structure available at each node,
whereas in DSDV, nodes only have distance estimates to other
nodes via a neighbor. PSR sits in the middle ground, where each
node maintains a spanning tree of the network. Furthermore,
DSR is a well-accepted reactive source routing scheme, and as
with PSR, it support source routing, which does not require
other nodes to maintain forwarding lookup tables. All three
baseline protocols are configured and tested out of the box
of ns-2.
In modeling node mobility of the simulated MANETs, we
use the random waypoint model to generate node trajectories. In
this model, each node moves toward a series of target positions.
The rate of velocity for each move is uniformly selected from
[0, vmax ]. Once it has reached a target position, it may pause
for a specific amount of time before moving toward the next
position. This mobility model may eventually lead to an uneven
node distribution in the network. That is, the nodes’ density
in the central area of the network may be higher than that at
the network boundary. This uneven node distribution coincides
with the real case in our daily life. However, at the beginning of
simulations, the nodes’ positions are evenly assigned; therefore,
we discard the simulation data in the first 30 s, and only the data
at a steady state is collected. All networks have 50 nodes in our
tests. We have two series of scenarios based on the mobility
model. The first series of scenarios have a fixed vmax but
different network densities by varying the network dimensions.
The second series have the same network density but varying vmax .
1 In our previous paper [15], PSR’s performance is also tested under a more
realistic physical model with opportunistic forwarding techniques.
Fig. 2. Routing overhead with density.
We study the data transportation capabilities of these routing
schemes and their overhead in doing so by loading the networks
with TCP data flows and UDP voice streams.
1) To test how TCP is supported, in each scenario, we randomly select 40 nodes out of the 50 and pair them up. For
each pair, we set up a permanent one-way File Transfer
Protocol (FTP) data transfer. We repeat the selection
of the 40 nodes five times and study their collective
behavior. Since TCP’s congestion avoidance mechanism
always tries to inject as much data in the network as
possible, this essentially mimics heavily loaded mobile
networks. For all four protocols, we measure their TCP
throughput, end-to-end delay, and routing overhead in
bytes per node per second in each scenario, where each
scenario has 20 simulation instances.
2) To study their performance in supporting UDP, we use
two-way constant-bit-rate (CBR) streams for compressed
voice communications. Specifically, we select three pairs
of nodes and feed each node with a CBR flow of
160 B/packets and 10 packets/s, which simulates mobile networks with a light voice communication load.
We measure the packet delivery ratio (PDR), end-to-end
delay, and delay jitter in each scenario.
Results about TCP (see Section IV-B and C) and UDP (see
Section IV-D and E) with regard to varying node densities and
velocity rates are in the following.
B. TCP With Node Density
We first study the performance of PSR, OLSR, DSDV,
and DSR in supporting 20 TCP flows in networks with different node densities. Specifically, with the default 250-m
transmission range in ns-2, we deploy our 50-node network
in a square space of varying side lengths that yield node
densities of approximately 5, 6, 7, . . . , 12 neighbors per node.
These nodes move following the random waypoint model with
vmax = 30 m/s.
We plot in Fig. 2 the per-node per-second routing overhead,
i.e., the amount of routing information transmitted by the
routing agents measured in B/node/s, of the four protocols when
they transport a large number of TCP flows. This figure shows
864
IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 63, NO. 2, FEBRUARY 2014
that the overhead of PSR (20 to 30) is just a fraction of that
of OLSR and DSDV (140 to 260) and more than an order of
magnitude smaller than DSR (420 to 830). The routing overhead of PSR, OLSR, and DSDV goes up gradually as the node
density increases. This is a typical behavior of proactive routing
protocols in MANETs. These protocols usually use a fixed-time
interval to schedule route exchanges. While the number of routing messages transmitted in the network is always constant for
a given network, the size of such message is determined by the
node density. That is, a node periodically transmits a message
to summarize changes as nodes have come into or gone out of
its range. As a result, when the node density is higher, a longer
update message is transmitted even if the rate of node motion
velocity is the same. Note that when the node density is really
high, e.g., around 10 and 12, the overhead of OLSR flattens
out or even slightly decreases. This is a feature of OLSR when
its multipoint relaying mechanism becomes more effective in
removing duplicate broadcasts, which is the most important
improvement of OLSR over conventional LS routing protocols.
PSR uses a highly concise design of messaging, allowing it
to have much smaller overhead than the baseline protocols.
In contrast, DSR, as a reactive routing protocol, incurs significantly higher overhead when transporting a large number of
TCP flows because every source node needs to conduct its own
route search. This is not surprising as reactive routing protocols
were not meant to be used in such scenarios. Later in our
experiments (see Section IV-D and E), we test all four protocols
supporting only a few UDP streams for a different perspective.
Here, the routing overhead of DSR decreases with the node
density going up and the network diameter going down. This
is because the number of hops to a destination is smaller in a
denser network; therefore, the shorter and more robust routes
break less frequently and do not need as many route searches.
Furthermore, compared with IP forwarding, the fact that DSR
is source routing and that intermediate nodes cannot modify the
routes embedded in data packets works against its performance
in a mobile network, both in terms of the increase in search
operations and the loss of data transportation capacity. The
reason is that, because a source node can be quite a few hops
away from the destination, its knowledge about the path as
embedded in the packets can become obsolete quickly in a
highly mobile network. As a packet progresses en route, if an
intermediate node cannot reach the next hop, as indicated in
the embedded path, it will be dropped. This is very different
from IP forwarding, where intermediate nodes can have more
updated routing information than the source and can utilize that
information in forwarding decisions.
Fig. 3 plots the TCP throughput of the four protocols for the
same node density levels as before. The total throughput of the
20 TCP flows of PSR, OLSR, and DSDV is noticeably higher
than that of DSR. In addition, while the TCP throughput of
DSR decreases with node density, that for the other three are
somewhat unaffected, hovering at around 500 kb/s. Apparently,
the large routing overhead of DSR, particularly in dense networks, consumes a fair amount of channel bandwidth, leaving
less room for data transportation. In most cases, PSR has the
highest throughput because it needs to give up the least network
resources for routing.
Fig. 3.
TCP throughput with density.
Fig. 4.
End-to-end delay in TCP with density.
Next, we focus on the end-to-end delay of TCP flows to
investigate how well these protocols support time-sensitive
applications. Fig. 4 shows the delay measured for different node
densities. As the density increases from 5 to 12 neighbors,
the delay of DSR goes up from 0.58 to about 1.5 s, which
is significantly higher than the typical value of 0.15 to 0.35 s
for the other three protocols. This difference is caused by
the initial route search when a TCP flow starts and by the
subsequent searches triggered by route errors. As the network
becomes denser, all protocols show an increasing trend in endto-end delay. This may seem counterintuitive as, in denser
networks, the average hop distance between source–destination
pairs is smaller, which should lead to shorter round-trip time.
However, this benefit is completely offset by more intense
channel contention. Recall that the node density is inversely
proportional to the square of the network diameter. As such,
in the interplay between route length and channel contention,
the latter dominates the overall effect.
C. TCP With Velocity
We also study the performance of PSR and compare it to
OLSR, DSDV, and DSR with different rates of node velocity.
In particular, we conduct another series of tests in networks
WANG et al.: PSR
Fig. 5.
Fig. 6.
Routing overhead with velocity.
TCP throughput with velocity.
865
Fig. 8. PDR in UDP with density.
to the right of the x-axis corresponds to the middle bars in
Fig. 2. We observe in the plot here that, as vmax decreases,
the overhead of all protocols comes down. The reason for DSR
is that, as the network structure becomes more stable, fewer
route repair attempts are necessary. For the case of the proactive
protocols, it is the reduction in the size of routing messages
(i.e., fewer neighbors have changed positions) that cuts down
the overhead. Still relative among these four protocols, when
the network is not stationary (vmax = 0), the overhead of PSR
(20–30 B/node/s) is a fraction of that of OLSR and DSDV
(90–300 B/node/s) and more than an order of magnitude lower
than DSR (180–770 B/node/s).
The TCP throughput and end-to-end delay are plotted in
Figs. 6 and 7, respectively. From these figures, we observe
that the performance of PSR, OLSR, and DSDV are similar
with PSR leading the pack in most cases. In addition, neither
throughput nor delay is affected by the different rates of velocity. The only exception is that, when vmax = 0, all protocols
yield a high throughput of 900 kb/s. With a greater portion
of the channel bandwidth devoured by routing messages in
highly mobile networks, DSR suffers a noticeable performance
penalty in TCP throughput and end-to-end delay.
D. UDP With Density
Fig. 7.
End-to-end delay in TCP with velocity.
of 50 nodes deployed in a 1100 × 1100 (m2 ) square area with
vmax set to 0, 4, 8, 12, . . . , 32 (m/s). The network thus has an
effective node density of around seven neighbors per node,
i.e., a medium density among those configured earlier. As with
before, 20 TCP one-way flows are deployed between 40 nodes,
and we measure the routing overhead, TCP throughput, and
end-to-end delay (see Figs. 5–7).
The routing overhead of all four protocols with varying rates
of node velocity is plotted as in Fig. 5. Note that the velocity
We also tested the four protocols for their performance in
transporting a small number of UDP streams. This is a typical
assumption for ideal scenarios of reactive routing protocols.
Here, we deploy three two-way UDP streams to simulate compressed voice communications. To find out about how node
density affects these protocols, we use the same network and
mobility configurations as in Section IV-B. We measure and
plot the PDR (see Fig. 8), delay (see Fig. 9), and delay jitter
(see Fig. 10) against varying node densities.
In Fig. 8, the PDRs of all four protocols are in the same
ball park across different node densities, with DSR slightly in
the lead and OLSR trailing behind. This verifies that the traffic
configuration is favorable for DSR. The relatively high loss rate
of OLSR among the proactive routing protocols is caused by the
higher routing overhead compared with PSR and DSDV. When
866
IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 63, NO. 2, FEBRUARY 2014
Fig. 9. End-to-end delay in UDP with density.
Fig. 11.
PDR in UDP with velocity.
Fig. 12.
End-to-end delay in UDP with velocity.
Fig. 10. End-to-end delay jitter in UDP with density.
E. UDP With Velocity
the nodes are neither too sparse so that the network connectivity
is good nor too dense so that the channel can be spatially reused,
these protocols have a fairly high PDR of over 70% for PSR,
DSDV, and DSR, and of 60%–70% for OLSR.
When we turn to end-to-end delay (see Fig. 9), there is a
noticeable difference between DSR and the proactive protocols.
In particular, DSR as a reactive protocol has a rather large delay
in sparse networks. This is because the long vulnerable routes
discovered during the search procedure break frequently, forcing nodes to hold packets back for an extended period before
new routes are identified. Conversely, the network sparsity does
not affect proactive protocols as much because their periodic
routing information exchange makes them more prepared for
network structure alteration. While the delay of DSR is off the
chart, that of PSR is always less than 0.05 s, which is also much
less than that for DSDV and OLSR (0.1–0.43 s). On a related
note, the delay jitter (see Fig. 10) of PSR is significantly lower
than the other three. Note that voice-over-IP (VoIP) applications
usually discard packets that arrive too late. Therefore, the
jitter among the packets actually used by the VoIP receiving
agent is much smaller. Nevertheless, our metric still reflects
how consistent these protocols are in delivering best-effort
packets.
The same measurements are taken to test these protocols in
response to different rates of node velocity. As with the case
earlier, we pick three node pairs out of the 50 nodes and give
them two-way CBR streams. For the entire series of different
velocity caps vmax = 0, 4, 8, 12, . . . , 32 m/s, the node density
is again set to around seven neighbors per node.
From the plot of PDR (see Fig. 11), we observe that DSR
is able to support three voice streams with little packet loss.
Specifically, the PDR of DSR, PSR, and DSDV is always
over 70% even when vmax = 32 m/s. The reliability of OLSR
is relatively lower, which can go below 60% at high speed
(vmax = 28 or 32 m/s). Note that all four protocols are very
reliable in data delivery when vmax = 0 or 4 m/s, where the
loss rates are well below 10%. Their performance in terms of
PDR degrades gracefully as the rate of node velocity increases.
The end-to-end delay (see Fig. 12) presents a rather distinct
landscape. In particular, the number for DSR is significantly
higher than the other protocols, except in low-mobility networks with vmax = 0 or 4 m/s. In all cases, the delay for PSR is
much smaller compared with OLSR and DSDV. On the other
hand, the measured delay jitter (see Fig. 13) indicates that
all protocols become less consistent when nodes move faster.
Relatively speaking, however, the variance of PSR is much
smaller than the other three.
WANG et al.: PSR
867
Fig. 13. End-to-end delay jitter in UDP with velocity.
V. C ONCLUSION
This paper has been motivated by the need to support
opportunistic data forwarding in MANETs. To generalize the
milestone work of ExOR for it to function in such networks, we
needed a PSR protocol. Such a protocol should provide more
topology information than DVs but must have significantly
smaller overhead than LS routing protocols; even the MPR
technique in OLSR would not suffice. Thus, we put forward a
tree-based routing protocol, i.e., PSR, which is inspired by the
PFA and the WRP. Its routing overhead per time unit per node is
on the order of the number of the nodes in the network as with
DSDV, but each node has the full-path information to reach all
other nodes. For it to have a very small footprint, PSR’s route
messaging is designed to be very concise. First, it uses only
one type of message, i.e., the periodic route update, both to
exchange routing information and as hello beacon messages.
Second, rather than packaging a set of discrete tree edges in
the routing messages, we package a converted binary tree to
reduce the size of the payload by about a half. Third, we
interleave full-dump messages with differential updates so that,
in relatively stable networks, the differential updates are much
shorter than the full-dump messages. To further reduce the size
of the differential updates, when a node maintains its routing
tree as the network changes, it tries to minimize alteration of the
tree. As a result, the routing overhead of PSR is only a fraction
or less compared with DSDV, OLSR, and DSR, as evidenced by
our experiments. Yet, it still has similar or better performance
in transporting TCP and UDP data flows in mobile networks of
different velocity rates and densities.
In the simulation in this paper, we used PSR to support
traditional IP forwarding for a closer comparison with DSDV
and OLSR, whereas DSR still carried source-routed messages.
In our simultaneous work, i.e., CORMAN [15], we tested PSR’s
capability in transporting source-routed packets for opportunistic data forwarding, where we also found that PSR’s small
overhead met our initial goal. That being said, as indicated
earlier in Section IV-B, while alleviating forwarding nodes from
table lookup, DSR’s source routing is particularly vulnerable
in rapidly changing networks. The reason for this is that, as
a source-routed packet progresses further from its source, the
path carried by the packet can become obsolete, forcing an
intermediate node that cannot find the next hop of the path to
drop the packet. This is fundamentally different from traditional
IP forwarding in proactive routing with more built-in adaptivity,
where the routing information maintained at nodes closer to
the destination is often more updated than the source node. Although out of the scope of this paper, it would be an interesting
exploration to allow intermediate nodes running DSR to modify
the path carried by a source-routed packet for it to use its more
updated knowledge to route data to the destination. This is in
fact exactly what PSR does when we used it to carry sourcerouted data in CORMAN. Granted, this opens up an array of security issues, which themselves are part of a vast research area.
As with many protocol designs, in many situations working
on PSR, we faced tradeoffs of sorts. Striking such balances not
only gave us the opportunity to think about our design twice
but also made us understand the problem at hand better. One
particular example is related to trading computational power
for data transfer performance. During one route exchange interval, a node receives a number of routing messages from its
neighbors. It needs to incorporate the updated information to its
knowledge base and share it with its neighbors. The question is
when should these two events happen. Although incorporating
multiple trees at one time is computationally more efficient, we
chose to do that immediately after receiving an update from a
neighbor. As such, the more accurate information takes effect
without any delay. Otherwise, when a data packet is forwarded
to a neighbor that no longer exists, it causes link layer retrial,
backlogging of subsequent packets, and TCP congestion avoidance and retransmission. With the broadcast and shared nature
of the wireless channel, the effects above are adversary to all
other data flows in the area. Therefore, in research on multihop
wireless networking, it usually makes sense for us to minimize
any impact on the network’s communication resources even if
there is penalty in other aspects. When it comes to the case
when a node should share its updated route information with
its neighbors, we chose to delay it until the end of the cycle
so that only one update is broadcast in each period. If a node
were to transmit it immediately when there is any change to its
routing tree, it would trigger an explosive chain reaction and
the network would be overwhelmed by the route updates. As
we found out in our preliminary tests, this is the primary reason
that WRP’s overhead was significantly higher than the other
protocols under study.
R EFERENCES
[1] I. Chlamtac, M. Conti, and J.-N. Liu, “Mobile ad hoc networking: Imperatives and challenges,” Ad Hoc Netw., vol. 1, no. 1, pp. 13–64, Jul. 2003.
[2] M. Al-Rabayah and R. Malaney, “A new scalable hybrid routing protocol
for VANETs,” IEEE Trans. Veh. Technol., vol. 61, no. 6, pp. 2625–2635,
Jul. 2012.
[3] R. Rajaraman, “Topology control and routing in ad hoc networks: A
survey,” ACM SIGACT News, vol. 33, no. 2, pp. 60–73, Jun. 2002.
[4] Y. P. Chen, J. Zhang, and I. Marsic, “Link-layer-and-above diversity
in multi-hop wireless networks,” IEEE Commun. Mag., vol. 47, no. 2,
pp. 118–124, Feb. 2009.
[5] P. Larsson, “Selection diversity forwarding in a multihop packet radio network with fading channel and capture,” ACM Mobile Comput. Commun.
Rev., vol. 5, no. 4, pp. 47–54, Oct. 2001.
[6] S. Biswas and R. Morris, “ExOR: Opportunistic multi-hop routing for
wireless networks,” in Proc. ACM Conf. SIGCOMM, Philadelphia, PA,
USA, Aug. 2005, pp. 133–144.
868
IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 63, NO. 2, FEBRUARY 2014
[7] T. Clausen and P. Jacquet, “Optimized Link State Routing Protocol
(OLSR),” RFC 3626, Oct. 2003. [Online]. Available: http://www.ietf.org/
rfc/rfc3626.txt
[8] D. B. Johnson, Y.-C. Hu, and D. A. Maltz, “On The Dynamic Source
Routing Protocol (DSR) for mobile ad hoc networks for IPv4,” RFC 4728,
Feb. 2007. [Online]. Available: http://www.ietf.org/rfc/rfc4728.txt
[9] C. E. Perkins and P. Bhagwat, “Highly dynamic Destination-Sequenced
Distance-Vector Routing (DSDV) for mobile computers,” Comput. Commun. Rev., vol. 24, pp. 234–244, Oct. 1994.
[10] C. E. Perkins and E. M. Royer, “Ad hoc On-Demand Distance Vector
(AODV) routing,” RFC 3561, Jul. 2003. [Online]. Available: http://www.
ietf.org/rfc/rfc3561.txt
[11] S. Murthy, “Routing in packet-switched networks using path-finding algorithms,” Ph.D. dissertation, Comput. Eng., Univ. California, Santa Cruz,
CA, USA, 1996.
[12] J. Behrens and J. J. Garcia-Luna-Aceves, “Distributed, scalable routing
based on link-state vectors,” in Proc. ACM Conf. SIGCOMM, 1994,
pp. 136–147.
[13] S. Murthy and J. J. Garcia-Luna-Aceves, “An efficient routing protocol
for wireless networks,” Mobile Netw. Appl., vol. 1, no. 2, pp. 183–197,
Oct. 1996.
[14] D. West, Introduction to Graph Theory, 2nd ed. Upper Saddle River, NJ,
USA: Prentice-Hall, Aug. 2000.
[15] Z. Wang, Y. Chen, and C. Li, “CORMAN: A novel cooperative opportunistic routing scheme in mobile ad hoc networks,” IEEE J. Sel. Areas
Commun., vol. 30, no. 2, pp. 289–296, Feb. 2012.
Zehua Wang (S’13) received the B.Eng. degree
from Wuhan University, Wuhan, China, in 2009 and
the M.Eng. degree from the Memorial University of
Newfoundland, St John’s, NL, Canada, in 2011. He
is currently working toward the Ph.D. degree with
The University of British Columbia, Vancouver, BC,
Canada.
His research interests include machine-type communication networking, wireless ad hoc networking,
mobile and distributed computing, and generic data
networks.
Mr. Wang served as a member of the Technical Program Committees for
the IEEE International Conference on Wireless and Mobile Computing, Networking, and Communications in 2011; the IEEE International Conference on
Communications (IEEE ICC), the IEEE Global Communications Conference,
and the IEEE International Conference on Connected Vehicles and Expo (IEEE
ICCVE) in 2012; and the IEEE ICCVE in 2013. He will serve as a member of
the Technical Program Committee for the IEEE ICCC in 2014.
Yuanzhu Chen (M’12) received the B.Sc. degree
from Peking University, Beijing, China, in 1999
and the Ph.D. degree from Simon Fraser University,
Burnaby, BC, Canada, in 2004.
From 2004 and 2005, he was a Postdoctoral
Researcher with Simon Fraser University. He is
currently an Associate Professor with the Department of Computer Science, Memorial University
of Newfoundland, St. John’s, NL, Canada. His research interests include computer networking, graph
theory, web information retrieval, and evolutionary
computation.
Cheng Li (SM’07) received the B.Eng. and M.Eng.
degrees from Harbin Institute of Technology, Harbin,
China, in 1992 and 1995, respectively, and the
Ph.D. degree in electrical and computer engineering
from the Memorial University of Newfoundland, St.
John’s, NL, Canada, in 2004.
He is currently a Full Professor with the Faculty of
Engineering and Applied Science, Memorial University of Newfoundland. His research interests include
mobile ad hoc and wireless sensor networks, wireless
communications and mobile computing, switching
and routing, and broadband communication networks.
Dr. Li is a registered Professional Engineer in Canada and a member of the
IEEE Communication, Computer, Vehicular Technology, and Ocean Engineering Societies. He has served as a Co-Chair of the Technical Program Committee
(TPC) of the IEEE Biennial Symposium on Communications in 2010 and the
IEEE International Conference on Wireless and Mobile Computing in 2011. He
has served as a Co-Chair for various technical symposia of many international
conferences, including the IEEE Global Communications Conference (IEEE
GLOBECOM) and the IEEE International Conference on Communications
(IEEE ICC). He has also served as a member of TPCs for many international
conferences, including the IEEE ICC, the IEEE GLOBECOM, and the IEEE
Wireless Communications and Networking Conference. He is an editorial
board member of Wiley Wireless Communications and Mobile Computing,
the Journal of Networks, the International Journal of E-Health and Medical
Communications, and the Korean Society for Internet Information Transactions
on Internet and Information Systems. He is also an Associate Editor for Wiley
Security and Communication Networks.

Similar documents

×

Report this document