|
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] Re: TCP (and SCTP) sucks on high speed networks> From: "Steven M. Bellovin" <smb@research.att.com> > To: Matt Wakeley <matt_wakeley@agilent.com> > Subject: Re: TCP (and SCTP) sucks on high speed networks > Date: Fri, 01 Dec 2000 17:14:05 -0500 ... > We, as a profession, don't know how to do better today, in the sense that > we don't have a solution that is (a) rationally deployable, (b) solves this > problem, (c) is "TCP-friendly", in the sense that it doesn't have undue > negative impact on existing RFC-compliant TCP streams, and (d) is > reasonably compliant with the Internet and the Internet philosophy. I'm not sure that I agree with that. It occurs to me that what is really at issue here is a question of how much the congestion window should be increased each RTT during CW avoidance. Or more specifically, the current algorithm (1 MSS) works well when the ratio between the delay-bandwidth (DBW) and the MSS is small, but as DBW gets larger, and the ratio get larger, the algorithm starts to break down. Since the CW gets pulled back back by a percentage value of the FlightSize (50%), perhaps it should also be increased by a percentage value of the FlightSize, rounded up to the nearest multiple of the MSS. Pick the right value, and it won't kick in until you get a severely lopsided ratio such as this example. BTW, what MSS is being assumed at 10Gbs? That will change the recovery time. If you have a 64 Kbyte MSS, you are going to recover a lot faster than with a 1500 byte MSS. Hence my idea that the discussion should focus on the size of the arithmetic increment. And I think a very focused discussion like that could produce some valid results in a fairly short time. -David Borman, dab@bsdi.com
Home Last updated: Tue Sep 04 01:06:12 2001 6315 messages in chronological order |