Fast TCP

1
Fast TCP

• Cheng Jin David Wei Steven Low
• Caltech
• Infocom, March 2004
• Offense Team Santa Animesh

2
Delay-based Approach

• In general, Equation-based approach is needed for
  steady-state and high utilization
• Congestion measure delay, loss,
• We agree with the argument Delay-based approach
  is better than loss-based approach
• Multi-bit information compared to one-bit
• Earlier detection of impending congestion
• More frequent feedback possible

3
Is the approach novel?

• NO. Delay-based approaches have been proposed for
  the last 15 years
• Including their earlier paper last year
• Half the theory of this paper is a copy of their
  earlier paper, IEEE CCW 2000

4
(Just) A few Prior Art

• Raj jain, A Delay-based approach for congestion
  avoidance in interconnected heterogeneous
  computer network. ACM Sigcomm, 89
• Martin, Nilsson and Rhee, Delay-based congestion
  avoidance for TCP, IEEE ToN, June 2003
• Sisalem and Schulzrinne, The Loss-Delay
  Adjustment Algorithm A TCP friendly adaptation
  Scheme. NOSSDAV, 98
• Rejaie, Handley, Estrin, RAPAn end-to-end
  Rate-based Congestion Control Mechanism for
  Realtime Streams in the Internet. Infocom 99
• Floyd et. al., Equation based congestion control
  for unicast applications. Sigcomm 00
• Brakmo and Peterson, TCP Vegas end-to-end
  congestion avoidance on a global internet. IEEE
  JSAC, Oct 1995
• Choe and Low, Stabilized Vegas. Infocom 03

5
Comparison with WHAT ???

• Compared their approach to 3 other loss-based
  approaches.
• Why not delay-based approaches
• Other approaches?
• Particularly, TCP-Vegas
• Explicit Congestion Notification (ECN)
• Sneaking suspicion Did they actually compare,
  and was on the losing side?

6
Inter-Protocol Fairness

• All transport protocols should be TCP friendly
  for compatible with a deployed standard
• TCP-friendliness Maintaining arrival rate to at
  most some constant over the square root of the
  packet loss rate
• No claim for TCP-friendliness in this paper

7
TCP-Friendly
8
Deployment Issues

• A proposal should have incremental deployment
  characteristic.
• Lot of excellent proposals have not seen the day
  of light because of this
• Examples TCP-Vegas, SACK, Multicast
• So, just yet another paperread forget

9
Evaluation Flaws

• Wrong comparison choices
• Incomplete evaluation

10
Intra-Protocol Fairness

• Intra-Protocol Fairness
• - Showed better fairness index but themselves
  claim it not being 1.
• - Compare against HSTCP, STCP which were
  protocols optimized for high throughput and not
  fairness
• - Why not compare against TCP WestWood and
  Easy RED or TCP-Vegas ?
• TCP WestWood and Easy RED to improve Fairness in
  high speed networks. PfHSN02

11
Intra-Protocol Fairness Contd.

• Maintaining intra-protocol fairness when some of
  the connections have large propagation delay is
  difficult.
• Global Fairness of Additive-Increase and
  Multiplicative-Decrease With Heterogeneous
  Round-Trip Times. Infocom 2000
• FAST-TCP fails here, but TCP Vegas was shown to
  successful.

12
Inter-Protocol Fairness

• TCP Vegas paper clearly demonstrates their
  effectiveness in being TCP friendly using
• When Reno competes with Vegas, Vegass flow does
  not throttle Reno.
• Also the total retransmissions drops when Reno
  competes with Vegas as compared to Reno against
  Vegas.
• Background trafficss throughput increased by 20
  when Reno is competing with Vegas as compared to
  Reno competing with itself.

13
Are you afraid to hit the roads ?

• Are conclusions based on dummynet experiments
  sufficient ?
• ns simulation, WAN emulation, real deployment

14
Conclusion

• Theoretically sound in advocating Delay Based
  approaches
• But this is prior work
• Did not convince me why FAST TCP is the choice as
  compared to other delay based approaches.
• Weak and Incomplete Evaluation
• Fairness issues is questionable
• Wrong comparison choices
• Experimental with realistic scenarios necessary

15
TCP-Vegas

• Expected Rate Curr wind Size / Min RTT
• Actual Rate Bytes sent during RTT / Current RTT
• Diff (Expected Actual) should remain within
  limits