TCP with Variance Control for
Multihop IEEE 802.11 Wireless
Networks
Jiwei Chen, Mario Gerla, Yeng-zhong Lee
How does TCP work?
• TCP protocol dynamically increases the
congestion window CWIN at each round
– However, if congestion is detected, the TCP source
reduces CWIN (by half, in TCP Reno)
– In most TCP varieties, congestion is “signalled” to
source by packet loss
• This strategy is effective in the Wired Internet:
– Congestion immediately causes buffer overflow and
packet loss
– The loss is quickly reported to source
Why packet loss not enough in
wireless nets?
• In wireless (802.11), we have ARQ:
– Packet is retx up to 5 times; timer doubled each time
– Only after 5 retx the packet is dropped and source
learns about it
– Packet drop comes too late!
– Pkt drop feedback latency leads to “oscillatory”
behavior, slow convergence, unfairness, capture etc
• Experimental result on wired and wireless links
in next 2 slides
CBR experiment: Throughput
Variance in Wired Network
Instantaneous CBR throughput vs. different rates on a wired link with
500 Kbps capacity
Rate Variance in Wireless Networks
Instantaneous CBR throughput vs. different rate on a 4 hop chain.
Channel data rate = 2Mbps
Why High Rate Variance in 802.11 nets?
• When congestion builds up (due to excessive
self rate or too many connections):
– Packets are dropped due to collisions and are
retransmitted
– Delay between pkts and ACKs fluctuates
– Observed data rate fluctuates
– And, data rate variance increases
– All this happens before any packet is lost!
• Rate variance more timely congestion indicator
than packet loss
Note: Variance increases also when packets are “randomly” dropped,
say, because of jamming - more later..
TCP-VAR
Combined: CWND Cong Control + Rate Control
• Congestion Control:
– Linearly Increase cwnd if throughput variance is
small
– Multiplicatively decrease cwnd if throughput
variance is large
– If packet loss detected, cut cwnd by half
• Rate control:
– Use pacing with rate given by:
cwnd
R=
var
srtt ´ e
Performance Evaluation
• Compare TCP-VAR, TCP adaptive pacing,
TCP small and TCP NewReno
• One or multiple TCP flow(s) over a chain
topology
• Channel bandwidth 2 Mbps
• Packet size 1460 bytes
• AODV routing is used.
Related Work
• TCP Optimal Window Concept
– Congestion Window Limit based on the round trip hop count
– Assumptions of the shortest path and ratio of
Interference/Transmission range
• Small window increase (TCP-Small)
• Congestion Window Limit based on the round trip hop count
(1 pkt every 4 hops..)
• The linear increase parameter is a small value, say 0.01, i.e
w=w+0.01/w (Mobihoc 05)
• TCP Adaptive Pacing (TCP-AP)
• 4 hop pacing (Mobihoc 05)
• Same assumptions as above.
One TCP on 4 hop Chain
Packet Losses
Protocol
Packet Loss
TCP NewReno
1655
TCP Small (0.01)
45
TCP AP
5
TCP VAR
0
Aggregate Throughput
1 TCP Flow on Varying Hop Length
Hop Count
Aggregate Throughput
4 TCP Flows on Varying Hop Length
Hop Count
Fairness and Efficiency
4 TCP flows on 4 hop chain
4 TCP flows on 6 hop chain
TCP-VAR keeps best fairness and efficiency
Conclusions
• TCP-VAR is purely end-to-end:
– independent of routing protocols,
– hop distance computation and
– ratio of Interference/Tx ratio.
• Novel congestion detection and congestion
control algorithm;
– It combines CWND and Rate adjustment
• TCP-VAR improves both efficiency and fairness.
Future Work
• Differentiate random loss and congestion loss
• Apply variance concept to other metrics
– Eg, RTT
• Friendliness issue
– Parameter adaptation to make it friendly to standard
TCP.
• Wired and wireless network integration