1. The Impact of Multihop Wireless Channel on TCP Performance
연구 목표
The Impact of Multihop Wireless Channel on TCP Performance
Z. Fu, H. Luo, P. Zerfos, S. Lu,
L. Zhang, and M. Gerla
IEEE Transactions on Mobile Computing, 2005
Presented by Chaegwon Lim

2. Summary Spatial channel reuse can improve channel utilization.
연구 목표
Summary
Spatial channel reuse can improve channel utilization.
A TCP window size W* (Hops/4) exists at which throughput is maximized by achieving best spatial reuse.
Standard TCP typically grows its average window much larger than W* in IEEE networks.

3. Link layer techniques to improve TCP throughput LRED
연구 목표
Link layer techniques to improve TCP throughput
LRED
Tune the wireless link’s drop probability to maintain CWND near W*.
Adaptive Pacing
Increase the spatial reuse of the channel

4. Contents Background Packet Loss Observation Optimal TCP Window Size W*
연구 목표
Contents
Background
Packet Loss Observation
Optimal TCP Window Size W*
TCP Performance in MWN
Proposal: LRED
Proposal: Adaptive Pacing
Comments
Conclusions

5. TCP and IEEE 802.11 DCF TCP IEEE 802.11 DCF
연구 목표
TCP and IEEE DCF
TCP
Transmission Control Protocol
Prevalent reliable transport protocol used in the Internet today.
IEEE DCF
De facto standard for wireless communication
The huge number of devices has been already deployed

6. TCP over MWN Multihop Wireless Networks have Lossy Channels
연구 목표
TCP over MWN
Multihop Wireless Networks have
Lossy Channels
Hidden\exposed terminal problems
Routing Failures
 TCP has been optimized for running only over wired networks. So some problems occurs

7. Hidden/Exposed Nodes A B C D E F G H I Hidden Node Problem A B C D E A
연구 목표
Hidden/Exposed Nodes
Carrier Sensing Range
Transmission Range A B C D E F G H I
Hidden Node Problem
Carrier Sensing Range of Node B
RTS A B C D E
CTS
Carrier Sensing Range of Node D
DATA
RTS A B C D E

8. A B C D E ExposedNode Problem A B C D E
연구 목표
Carrier Sensing Range of Node D
Data A B C D E
ExposedNode Problem
Carrier Sensing Range of Node B
RTS
Data A B C D E

9. Two Causes of Packet Loss
연구 목표
Two Causes of Packet Loss
Sender: successive retransmissions
Receiver:
Buffer overflow

10. Packet loss Link-layer contention happens before the buffer overflow.
연구 목표
Packet loss
Link-layer contention happens before the buffer overflow.
If buffer > 20 pkts
 Buffer overflow is never observed in most cases.
 Link-layer contention is a major source of packet losses
The gradually increasing packet dropping probability due to link-layer contention is insufficient to stabilize the TCP window size around optimal value.

11. Packet Loss Observations
연구 목표
Packet Loss Observations
- Link Drop is dominant.
- No buffer overflow
if the buffer size is more than 20

12. Spatial Reuse Nodes are 200m apart Transmission range is 250m
연구 목표
Spatial Reuse
Nodes are 200m apart
Transmission range is 250m
Carrier sensing and interference range is 550m
pkts can be delivered at a same time
Spatial Reuse can improve the network throughput 2

13. TCP Window Size in Chain Topology
연구 목표
TCP Window Size in Chain Topology
Max concurrent senders are h/4, where max spatial reuse is achieved
TCP window size < h/4
 under utilization
TCP window size > h/4
 reduced throughput
due to packet losses and degraded spatial reuse.
(hidden/exposed problem)
The chain topology represents the packet forwarding
path generated by a minimum-hop routing protocol
such as DSR and AODV.

14. Simulation 7 hop chain TCP window = min (CWND, MaxCWND)
연구 목표
Simulation 7 hop chain
TCP window = min (CWND, MaxCWND)
W* and h/4 match reasonably well.
Really?, but maximum is three not two
TCP packets in flight do not distribute evenly among nodes

15. TCP Optimal window size in chain topologies of different lengths.
연구 목표
TCP Optimal window size in chain topologies of different lengths.
Chain topology
(1 flow, 7 hop)

16. Cross Topology 2 TCP flows Best window W* = 2,  measured window = 12
연구 목표
Cross Topology
2 TCP flows
Best window W* = 2,  measured window = 12
20% throughput reduction

17. Grid Topology 4, 8, and 12 TCP flows ½ of flows in each direction
연구 목표
Grid Topology
4, 8, and 12 TCP flows
½ of flows in each direction
Measured TCP windows are larger than optimal WND

18. 연구 목표
Simulation Results
Max Throughput and Win Size

19. Why TCP throughput decreases at CWND > W*?
연구 목표
Why TCP throughput decreases at CWND > W*?
# of competing nodes increases
 level of link-layer contention increases
 hidden/exposed node scenarios happens more
 many collisions and packet loss.
Large TCP CWND contributes the more packets in flight
 increasing # of competing node.

20. TCP Packet Drop Probability
연구 목표
TCP Packet Drop Probability
One flow, 7 hop chain
PDP is small at CWND is 2, so CWND increases beyond W*.

21. UDP Packet Drop Probability
연구 목표
UDP Packet Drop Probability
PDP is Saturated, when all nodes have backlogged packets  Channel Capacity

22. Analysis of Link Drop Probability
연구 목표
Analysis of Link Drop Probability
Modeling a random topology, drop probability is
m: # of competing nodes
c(m): # of nodes which transmit RTS successfully
b(m): # of nodes which transmit DATA successfully

23. Three regions of behavior
연구 목표
Three regions of behavior
Pl ~0: m, number of backlogged nodes, is < B*, maximum number of concurrent DATA transmitting nodes, and m~b~c
Pl increases linearly: m>B* and m<C*, maximum number of nodes with a clear channel
Pl stable: m>C* - the amount of contention cannot increase

24. A B C D E A B C D E Carrier Sensing Range of Node D Data
연구 목표
Carrier Sensing Range of Node D
Data A B C D E
Carrier Sensing Range of Node B
RTS
Data A B C D E

25. Motivation of Link Layer RED
연구 목표
Motivation of Link Layer RED
Random Early Drop
Link Drop
When a network is overloaded, the link drop probability starts to increase.
As the network load further increases, then link drop probability starts to saturate.
RED  By dropping before the queue (channel capacity) is full, the throughput of TCP can be improved.

26. LRED Link layer maintains the average number of retries.
연구 목표
LRED
Link layer maintains the average number of retries.
Next packet is dropped/marked with probability based on the average number.
If the average number of retries is small, packets are not dropped/marked.

27. 연구 목표
Motivation of Pacing
Balancing traffic among nodes can improve spatial channel reuse.
Let a node backoff an additional packet transmission time if the traffic load is high.

28. Adaptive Pacing Enabled from LRED
연구 목표
Adaptive Pacing
Enabled from LRED
If average retries < min_th calculate backoff time as usual
If pacing, backoff time increases by a time equal to the transmission time of the previous packet

29. Performance Improvement
연구 목표
Performance Improvement
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
LRED
Pacing

30. TCP Performance Chain Topology
연구 목표
TCP Performance Chain Topology

31. 연구 목표
Multiple TCPs

32. Average TCP Window Size
연구 목표
Average TCP Window Size

33. Conclusions Spatial channel reuse can improve channel utilization.
연구 목표
Conclusions
Spatial channel reuse can improve channel utilization.
A TCP window size W* (Hops/4) exists at which throughput is maximized by achieving best spatial reuse.
Standard TCP typically grows its average window much larger than W* in IEEE networks.
Link layer techniques to improve TCP throughput
LRED
Tune the wireless link’s drop probability to maintain CWND near W*.
Adaptive Pacing
Increase the spatial reuse of the channel

34. Comments This paper ignores the packet loss due to routing breakage.
연구 목표
Comments
This paper ignores the packet loss due to routing breakage.
How can we calculate W* for flows in complex topologies? It is not clear.
The analysis considers only the exposed node problem and does not includes the hidden node problem.
Mathematical reasoning for LRED is not enough.

35. 연구 목표
Question?