Increasing TCP`s Initial Window

Document technical information

Format pdf
Size 579.4 kB
First found May 22, 2018

Document content analysis

Category Also themed
Language
English
Type
not defined
Concepts
no text concepts found

Persons

Organizations

Places

Transcript

Increasing TCP initial window
draft-hkchu-tcpm-initcwnd-01.txt
Nandita Dukkipati
Yuchung Cheng
Jerry Chu
Matt Mathis
{nanditad, ycheng, hkchu, mattmathis}@google.com
30 July, 2010
78th IETF, Maastricht
Overview of prior results for IW10
Our proposal: increase TCP IW to 10 MSS
IW10 improves average TCP latency by ~10%
Large scale data-center experiments demonstrate latency
improves across network and traffic properties:
Varying network BW, RTTs, BDP, HTTP response
sizes, mobile networks
Small overall increase in retransmission rate (~0.5%),
with most from multiple connections
Prior work:
http://www.ietf.org/proceedings/10mar/slides/tcpm-4.pdf
http://ccr.sigcomm.org/online/?q=node/621
New contributions and the questions
addressed
A framework for running experiments with different IWs in
the same data-center
Primary concern from IETF-77: how does IW10 perform on
highly multiplexed links such as in Africa and South
America?
What is the impact on latency due to losses in IW?
Evaluated the impact of different IWs [3, 10, 16] on latency
and retransmission rate
Reinforced the prior experiment results with IW10
Testbed experiments for IW study in controlled environment
Preliminary results on fairness
Improved methodology for experiments
Previous methodology:
Change IW for entire data-center every week
Less apples-to-apples: changes in server software and
user base
Takes weeks to collect data
New methodology:
Serve different IWs based on IP address in one data-center
simultaneously for weeks
Same IW for connections from the same IP
More apples-to-apples: similar load across server
software update and user churn
Analysis of IW10 on Africa traffic
Experiment for 1 week in
June 2010
Impact of IW10 on Africa traffic
Web search latency (ms) and retransmission rate %
All of Africa
Percentile
Avg.
50
75
90
99
IW=10
988.4
503
795
1467
5042
IW=3
1123.9
538
878
1710
5923
IW=10
3.77%
Impr.
135.5
35
83
243
881
IW=3
3.35%
% Impr.
12%
6.5%
9.5%
14.2%
14.9%
Increase
0.42
Retrans. %
Africa with low QPS
Percentile
Avg.
50
75
90
99
IW=10
1870.5
733
1363
3146
11579
IW=3
2340.7
857
1773
4110
14414
IW=10
6.71%
Impr.
470.2
124
410
964
2835
IW=3
5.83%
14.5%
23.1%
23.5%
19.7%
Increase
0.87
% Impr. 20.1%
Retrans. %
Why does latency improve in Africa?
Large network round-trip time
Larger IW helps faster recovery of packet losses
Experiments on testbed demonstrate latency improves in
spite of increased packet losses
Why does latency improve in Africa?
Testbed experiment: 20Mbps, RTT 300ms, BDP buffer, offered
load 0.95, 50KB response size
Motivating example: Makerere University, Uganda
Completion time
Retransmission rate
Analysis of IW10 on South America traffic
Experiment for 1 week in
June 2010
Latency improvement across services in South
America
Latency improves across a variety of services
Services with multiple connections experience:
Least latency benefits
Most increase in retransmission rate
Percentile
Web
iGoogle
News
Blogger Photos
(multiple
connections)
10
18 [6%]
30 [10%]
4 [2.5%]
2 [1.1%]
6 [3.8%]
50
38 [6.6%]
198 [26%]
45 [9.9%]
98 [12.7%]
12 [3.2%]
90
154 [11%]
430 [16%]
336 [15%]
251 [4.5%]
37 [2.6%]
99
561 [12%] 986 [9.7%] 1827 [19%]
691 [2.9%]
134 [2.9%]
2.93
1.28
Delta in
Retrans %
0.51
0.52
entry: latency improvement (ms) [% improvement]
0.35
Maps
(multiple
connections)
Impact of latency under packet losses
Latency of traffic with retransmissions > 0 improves with IW10 as
compared to IW3
% traffic with rexmit > 0
IW3
IW10
All
6.6%
6.8%
Web
Search
6.11%
6.57%
Retransmissions of IW3 vs IW10
IW10 has ~0 increase in
#timeouts, but has more
fast-retransmit
post-RTO retransmits
Experiments with higher IWs
Does IW > 10 show better latency?
Try IW = {3, 10, 16} at
DC 1
20% in US east coast (RTT < 100ms)
80% in south America (RTT > 100ms)
DC 2
97% in Europe (RTT < 100ms)
Comparison of IW = 3, 10, 16 (DC 1)
Small improvement for larger IWs (>10); mostly for mid-size flows
static-content
Small/Mid:
0-64KB
Large:
>64KB
Comparison of IW = 3, 10, 16 (DC 2)
Small improvement for larger IWs (>10); mostly for mid-size flows
static-content
Testbed topology
All results are preliminary!
1GE
10GE
switch
1GE
10GE
Linux box as a router
with netem to emulate
N buffers, 300ms RTT
20Mbps bottleneck
switch
Traffic generator – enhanced netperf dispatched based on poisson arrival
Offered load - # of conn/sec (λ) with fixed response size, no pipelining
Tests parameters - bottleneck b/w, RTT, buffer space, response size
Test metrics - user completion time (UCT), retransmission rate, link
utilization
Measurement & Diagnosis tools
Fairness between IW10 and IW3 flows
50/50 mix of IW3 and IW10 traffic
BDP buffer, load 0.95, 15KB response size
Fairness between IW10 and IW3 flows
Same as previous slide except response size is 50KB
Conclusion
Take away summary
IW10 improves latency even in Africa and South
America
IW10 helps in quicker recovery from packet losses
A higher retransmission rate does not necessarily
increase latency
IW16 shows marginal latency improvement over IW10
Next steps
Ongoing work: fairness between IW3 and IW10 in the
transition phase
For any pending issues with IW10, join us in solving the
problems!
Steps to configure IW on Linux
Changing TCP IW on Linux (kernel version >= 2.6.30)
On your server, do
$ ip route show
select the outgoing route then do
$ ip route change default via <gateway> dev eth0 initcwnd <iw>
If the server process explicitly set SNDBUF, then SNDBUF value >= IW*MSS. Otherwise increase
the initial socket buffer if IW*MSS > /proc/sys/net/ipv4/tcp_wmem[1]
$ cat /proc/sys/net/ipv4/tcp_wmem
4096 16384 4194304
$ echo '4096 IW*MSS 4194304' > /proc/sys/net/ipv4/tcp_wmem
Must restart server process to use new tcp_wmem[1]
Acknowledgements
We acknowledge the following people at Google for their
contributions towards the large scale IW experiments:
Ethan Solomita
Elliott Karpilovsky
John Reese
Yaogong Wang
Roberto Peon
Arvind Jain
Why does latency improve in Africa?
(from tesbed experiment results)
Utilization
Timeouts
Fast Retransmits
Spurious
Timeouts
More preliminary results from testbed:
latency improves across all transaction
sizes
with BDP buffer & < 90% offered load
But retransmission rates can be quite
different
with BDP buffer and > 95% offered load
Insufficient buffer can hurt IW10 latency
40% BDP buffer, 75% offered load
×

Report this document