[e2e] 10% packet loss stops TCP flow

[Kostas Pentikousis]

On Fri, 25 Feb 2005, Roy Xu wrote:

|it seems to be a common understanding that if a TCP flow experiences
|10% or more packet loss, the flow stops (i.e., attains 0 or meaningless throughput)

Calculating drop rates from packet loss and using that to perform
Bernoulli trials is a poor error model for TCP.  First, summarily
counting how many packets were sent and how many of them were
dropped and calculating a dropped rate (say, 10%) captures very
little of what happens in reality. And it's particularly
inadequate if you go one more step and use that drop rate in an
ns2 simulation study. Second, the TCP retransmission rate (the
percentage of packets that are retransmitted) cannot always be
equated to packet loss rate. And it is the retransmission rate
that really matters if you are interested in high TCP throughput.

Nevertheless, a rate-based 10% (drop 1 in 10 packets coming by)
does not halt TCP -- at least it didn't in our lab tests:

K. Pentikousis and H. Badr, "Error modeling for TCP simulations",
Proceedings of EUROSIM 2001, Delft, The Netherlands, June 2001,
see esp. Fig. 2 (http://www.cs.stonybrook.edu/~kostas/art/)

And yes, that was TCP Tahoe and x-kernel ;)

|ns-2 simulations also seem to agree with this observation.

Last time I checked, the error models in ns2 do NOT drop pure
ACKs. Keep that in mind esp. for one-way TCP simulations. Not sure
if FullTCP is any different.

Best regards,

Kostas
__________________________________________________________________
Kostas Pentikousis                   www.cs.stonybrook.edu/~kostas