Scalable TCP: Improving Performance in Highspeed Wide Area Networks

1
Scalable TCP Improving Performance in Highspeed
Wide Area Networks

• Paper by Tom Kelly
• Presented by
• Christoph Jechlitschek

2
The Problem

• TCP performs badly with very large congestion
  windows
• Converges slowly
• Underutilizes available link capacity
• Example A 1 Gbps flow needs 28 minutes to
  recover from a single packet loss

3
TCP congestion window update

• cwnd ? cwnd 1/cwnd
• - if no loss was detected
• cwnd ? cwnd/2
• - if a loss was detected

4
TCP congestion window
5
Scalable TCP congestionwindow update

• cwnd ? cwnd 0.01
• - if no loss was detected
• cwnd ? cwnd 0.125 cwnd
• - if a loss was detected

6
Scalable TCP congestion window
7
Analysis

• Generalized form of update function
• cwnd ? cwnd a if no loss
  detected
• cwnd ? cwnd b cwnd if loss detected
• What makes a 0.01 and b 0.125 a good choice?

8
Legacy connections

• Traditional connections
• - are not designed to use large windows
• - limited amount of buffers
• Legacy window size lwnd 16
• Legacy loss rate pl 5.86 10-3

9
Response curve

• Relates a and b
• Choosing one fixes the other

10
Instantaneous rate variation

• Avoid large instantaneous rate variations
• Coefficient of variance depends on b only
• Therefore b should be as small as possible

11
Convergence

• Converge to a new bound as fast as possible
• Large a and b speed up convergence
• Contradicts with earlier statement

12
Decision

• Balance variation with converge time
• Choose a 0.01 and b 0.125

13
Experiment setup
14
Gigabit kernel modifications

• Increases size of kernel interface queues
• Avoids unnecessary packet copying
• Removes debug counters

15
Performance test 1

• 4 sender/receiver pairs transfer 2 GB files
• Count number of transfers in 1200 seconds

16
Performance test 1 results
17
Performance test 2

• Simulate Web traffic using standard TCP
• Background bulk transfer with Scalable TCP

18
Performance test 2 results
19
Conclusion

• Simple sender-side modifications greatly improved
  performance
• No modification to receiver needed
• Negligible effect on standard TCP connections
• Incrementally deployable

20
Questions?