The Impact of Multihop Wireless Channel on TCP Throughput and Loss

1
The Impact of Multihop Wireless Channel on TCP
Throughput and Loss

• Zhenghua Fu, Petros Zerfos, Haiyun Luo, Songwu
  Lu, Lixia Zhang, Mario Gerla
• INFOCOM2003, San Francisco, April 2003
• Presented by Philip Hardebeck

2
Outline

• Introduction
• Background
• TCP Throughput
• Several Topologies Chain, Cross, Grid, Random
• Simulations, Experiments, Analysis
• Proposed Solutions
• Conclusions

3
Introduction

• Do TCP mechanisms work well for Wireless Multihop
  Networks (WMN)?
• WMNs differ from wired networks.
• There is an optimal TCP window size for a given
  topology and flow pattern.

4
More Introduction

• Packet losses increase as window size exceeds
  optimal, up to a threshold.
• Link-RED and Adaptive Pacing are proposed to
  increase throughput.

5
Background MAC Basics
RTS
A
B
C
D
E
CTS
CTS
DATA
A
B
C
D
E
ACK
ACK
RTS
RTS
RTS
8 x
A
B
C
D
E
random exponential backoff ...
RTS
A
B
C
D
E
CTS
6
Spatial Reuse and Contention
Interfering Range
Communication Range
A
B
C
D
E
F
G
H
I
Interfering/Carrier Range of Node B
RTS
A
B
C
D
E
CTS
Interfering/Carrier Range of Node D
RTS
DATA
A
B
C
D
E
7
TCP Throughput

• Look at TCP throughput to show how well or poorly
  it performs spatial reuse.
• Typical TCP operation doesnt do a good job and
  the throughput is reduced.
• Identify window size for highest throughput, and
  verify with hardware experiments.

8
Chain Topology

• Packets of a single flow interfere with one
  another.
• Optimal window size is 1/4 number of hops in
  the chain.

9
Optimal Window Size vs. Chain Length
10
Throughput for 3 Packet Sizes
11
Actual vs. Simulated Throughput
12
Cross and Grid Topologies
13
Aggregate Throughput and Window Size
Table III
14
Throughput Summary

• Optimal window size exists for all topologies and
  flow patterns.
• Optimal window size derivable only for simple
  configuration (chain).
• Average TCP window size is much larger than
  optimal
• Causes more packet drops and reduced throughput

15
Loss Behavior

• Buffer drop probability is not significant in
  WMN, but contention drop is.
• Network overload is no longer a bottleneck link
  property, but a shared feature of multiple
  links.
• Drop probability increases gracefully as load
  increases.

16
TCP Packet Drop Probability
17
UDP Packet Drop Probability at MAC layer
18
Contrasting Drop Characteristics
19
Analysis of Link Drop Probability

• Modeling a random topology, drop probability is
• Three regions of behavior
• Pl 0 m, number of backlogged nodes, is lt B,
  maximum number of concurrent DATA transmitting
  nodes, and mbc

20
Analysis of Link Drop Probability Continued

• Other two regions
• Pl increases linearly mgtB and mltC, maximum
  number of nodes with a clear channel
• Pl stable mgtC - the amount of contention cannot
  increase

21
Link-RED Algorithm
22
Adaptive Pacing Algorithm
23
TCP Throughput Comparison
24
Multiflow TCP Throughput Comparison
25
Average TCP Window Size Comparison
26
Discussions

• TCP Vegas doesnt work as well as New Reno.
• Optimal window sizes exist for flows with
  variable packet size, but more complicated.
• LRED and Adaptive Pacing improve drop behavior.

27
Related Work

• Link-layer retransmission hides channel errors
  from upper layers
• Dynamic ad hoc networks and link failure are
  studied (routing issues)
• Studies of TCP ACK traffic using other MAC
  protocols
• Capacity of ad hoc networks using UDP/CBR flows

28
Conclusions

• TCP throughput improves if the window size
  operates at optimal, maximizing channel spatial
  reuse.
• TCP typically operates with a much larger window,
  reducing throughput.
• Wireless nodes exhibit a graceful drop feature.
• LRED and Adaptive Pacing improve throughput by up
  to 30

29
Problems/Weaknesses

• No explanation for the 10 difference between
  simulation and experimental results.
• Use of aggregate rate and window size makes it
  difficult to compare results to other papers.

30
Acknowledgements

• Thanks to Professor Kinicki for the opportunity
  to make this presentation.
• Thanks to Shugong Xu and Tarek Saadawi of CUNY
  for the MAC Basics and Spatial Reuse and
  Contention graphics.

31
Questions/Comments?