Limited Slow Start

1
Limited Slow Start

• Roman AroraJustin Roberts

2
Introduction

• Current Problem
• Experimental Solution
• Comparison
• Slow-Start vs. Limited Slow-Start
• Conclusion
• QA

3
Why Bother?

• Performance
• Quality of Service

4
Current Problem

• Internet nature has changed
• HTTP instead of Telnet
• Slow-Start actually very aggressive
• Exponential growth
• When loss occurs, threshold set at half current
  congestion window, and algorithm restarts
• When threshold reached, increase linearly

5
Current Slow Start
6
Slow Start Comparison
7
Current Problem

• Slow-Start can be manipulated by current TCP
  congestion avoidance algorithms
• TCP Reno
• TCP Tahoe
• TCP Vegas
• These algorithms further modify the congestion
  window value
• Need a more gentle adjustment algorithm once we
  know the approximate bandwidth

8
Experimental Solution

• Limited Slow Start
• Good for large congestion windows
• Limited growth instead of exponential growth
• Theory behind Reducing dropped packets by
  limited cwnd growth yields better performance
  than exponential cwnd growth.

9
Slow Start ? cwnd vs. time
cwnd
ss_thresh inf.
ss_thresh
ss_thresh
Timeout
Timeout
AIMD
AIMD
exp
exp
Slow Start
Slow Start
Slow Start
Time
10
Algorithm Comparison

• Slow Start
• Initially
• cwnd MSS
• ss_thresh inf.
• New ack received
• if(cwnd lt ss_tresh)
• / Slow Start /
• cwnd MSS
• else
• / Congestion Avoidance /
• cwnd MSS / cwnd
• Timeout
• / MD Multiplicative Decrease /
• ss_thresh cwnd / 2
• cwnd MSS

• Limited Slow Start
• Initially
• cwnd MSS ss_thresh inf
• max_ssthresh 100 MSS
• New ack received
• if(cwnd lt ss_tresh)
• if(cwnd lt max_ssthresh)
• cwnd MSS
• else
• / Limited Slow Start /
• k cwnd / (0.5 max_ssthresh)
• cwnd (MSS/k)
• else
• / Congestion Avoidance /
• cwnd MSS / cwnd
• Timeout
• / MD Multiplicative Decrease /

11
Limited Slow Start ? cwnd vs. time
cwnd
ss_thresh
ss_thresh inf.
ss_thresh
Timeout
AIMD
AIMD
Max_ssthresh
exp
exp
Slow Start
Slow Start
Slow Start
Time

• Limited Slow start introduces a new variable
  max_sstresh

• Limited Slow start increase rate is (1/K)MSS

12
Numerical Example

• It takes 16 round trips with slow start to reach
  a cwnd of 83,000 packets.
• With max_ssthresh (100 MSS), It takes 836
  round trips to reach a cwnd of 83,000 packets in
  Limited Slow Start.

13
Should it really work?

• ns/2 simulated path (receiver max window 1000
  bytes)

• It took slow start 1.5s reach max cwnd, while
  limited slow start took 3.75s

14
Experimental Results
15
So what does this mean?

• If you have a large congestion window.
• If you are not really being limited by the
  receiver window
• (rwnd gt swnd gt cwnd gt ss_thresh)
• THEN
• By limiting the maximum increase in the
  congestion window in a round-trip time, Limited
  Slow-Start can reduce the number of drops during
  slow-start, and improve performance of TCP
  connections

16
Different max_ssthresh values

• Why use (100 MSS) for the max max_sstresh
  variable?

17
Is it really that important?!

• Benefits network performance for large congestion
  windows
• Generally multiple connections from single user
  eases the slow start problem
• Not a problem for most scenarios
• Other mechanisms are available

18
Conclusion

• Why isnt it used?
• Currently experimental
• Only proposed in March 2004
• Part of HSTCP study
• What can be better?
• Improve Fast Retransmit algorithm to recover from
  multiple packet loss
• Rate-based pacing
• Appropriate Byte Counting

19
References

• http//www.wikipedia.org
• http//www.csm.ornl.gov/dunigan/net100/floyd.html
  
• http//www.acm.org/sigs/sigcomm/sigcomm96/papers/h
  oe.pdf
• http//www.eecs.umich.edu/zmao/eecs489/LectureSli
  des/congestion_control_6.pdf

20
Any Questions?