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 (UCLA), INFOCOM
2003, San Francisco, Mar. 2003.

• Presented by
• Scott McLaren

2
Overview

• Introduction
• Background
• Throughput in Multihop Wireless Networks
• Loss Behavior
• Improving Performance
• Conclusions

3
Introduction

• Improve channel utilization by spatial channel
  reuse
• A TCP window size W exists at which throughput
  is maximized by achieving best spatial reuse
• Increasing the window size past W will reduce
  throughput
• Standard TCP typically grows its average window
  much larger than W

4
Techniques to improve efficiency

• Link-RED
• Tune the wireless links drop probability
• Adaptive link-layer pacing scheme
• Increase the spatial reuse of the channel
• Allow TCP to operate in the contention avoidance
  region

5
802.11

• RTS/CTS messages
• Nodes hearing this handshake defer transmission
  until current transmission is finished
• Data is dropped if no CTS is received after 7 RTS
  retries
• Data is also dropped if 4 transmissions are sent
  without an receiving an ACK

6
Hidden Terminals

• A hidden terminal is a node in the receivers
  neighborhood, that cant detect sender and may
  disrupt transmissions

7

• Nodes are 200m apart
• Transmission range is 250m
• Carrier sensing and interference range is 550m
• D is a hidden terminal of A ? B
• Cannot hear CTS ( gt 250m )
• Cannot hear data from A, A is outside of Ds
  carrier sensing range
• D can transmit to E
• Causes collision at B, since D is within 550m
  interference range for B
• Contention loss at B

8
Chain Topology

• Best throughput when window size is h/4
• Assuming ideal MAC protocol and equal packet
  sizes
• Max concurrent senders is h/4, where max spatial
  reuse is achieved
• TCP window size lt h/4 ? under utilization
• TCP window size gt h/4 ? reduced throughput

9
(No Transcript)
10
Cross Topology

• 2 TCP flows
• Best window W 2, measured window 12
• 20 throughput reduction

11
Grid Topology

• 4, 8, and 12 TCP flows
• ½ of flows in each direction
• Measured TCP windows are larger than max
  achievable throughput

12
Results
13
TCP Loss Behavior

• Using 8-hop chain, all 165 TCP drops out of 12349
  transmissions were due to link drops

14
TCP Loss Behavior
15
Corollaries

• m number of backlogged nodes
• B the max number of nodes that can transmit
  their DATA packets concurrently without collision
• C denotes the max number of nodes that can
  initiate RTS messages
• Corollary 4.1
• m lt B
• Pl 0
• Corollary 4.2
• m gt B
• Pl increases as m increases
• Corollary 4.3
• m gt C
• Pl remains constant
• Throughput reduction due to Wavg gtgt W, Pl gt 0,
  Link contention gt 0 reducing spatial reuse

16
Improving TCP Performance

• Distributed Link RED (LRED)
• Adaptive Pacing

17
LRED

• Easy way is to improve performance by reducing
  buffer size, but problems with bursty traffic
• LRED exploits dropping in 802.11 MAC
• RED provides a linearly increasing drop curve as
  queue exceeds a min size
• LRED provides a linearly increasing drop curve as
  link drop probability exceeds a min size

18
LRED

• Link layer maintains average number of retries
• Next packet is dropped/marked with probability
  based on average number
• If average number of retries is small, packets
  are not dropped/marked
• When retries increase, the dropped/marked
  probability is calculated

19
Adaptive Pacing

• Improve spatial channel reuse by balancing
  traffic among nodes
• Exposed receiver problem
• Let a node backoff an additional packet
  transmission time when necessary

20
Adaptive Pacing

• Enabled from LRED
• If average retries lt min_th calculate backoff
  time as usual
• If pacing, backoff time increases by a time equal
  to the transmission time of the previous packet

21
Performance

• Chain Topology
• In all cases LRED Pacing increased TCP
  throughput by up to 30
• TCP stabilizes at a window size close to the
  optimal value
• The longer the chain, the better the improvement,
  due to pacing optimizing spatial channel reuse

22
Chain Topology
23
Performance

• Cross Topology
• Increased throughput and improves fairness
  (Jains) for both flows
• TCP NewReno has large unfairness, due to 802.11
  capture characteristic (collision of 2 packets,
  one weaker than the other. The stronger packet
  is received)

24
Performance

• Grid Topology
• Also increases throughput and fairness

25
Conclusions

• Only when buffer is small do buffer overflow
  drops dominate
• As buffer increases, link-layer drops dominate
• Link drop acts as a RED gateway, but is
  insufficient

26
Questions