Date: Wed, 10 Mar 2004 14:38:40 -0500 From: Mark Allman <mallman@icir.org> To: Brooks Davis <brooks@one-eyed-alien.net> Cc: Petri Helenius <pete@he.iki.fi> Subject: Re: Who wants SACK? (Re: was My planned work on networking stack) Message-ID: <20040310193840.6479F77A6D4@guns.icir.org> In-Reply-To: <20040310192255.GD14892@Odin.AC.HMC.Edu>
next in thread | previous in thread | raw e-mail | index | archive | help
--=-=-= > I looked for the paper I paraphrased, I'm pretty sure if was one by > Sally Floyd. I don't have the paper reference handy, but it's Sally's HighSpeed TCP work. I do happen to have a blurb on it sitting here that I think captures it well... Think of a network with an RTT of 100ms, a 1500 byte packet size and 10Gbps capacity. That means the congestion window needs to be 83,333 packets to fill the pipe (just by looking at the delay*bandwidth product of the network). And, to sustain this rate you need at most one loss every 5,000,000,000 packets (from the TCP model). That translates into about one loss every 100 minutes. And, that seems like a fairly large stretch. (This actually might be laid out in RFC3649.) However, this is a bit off-topic from SACK. Because this is all based on the AIMD nature of TCP's congestion control, not really on whether you employ SACK. (The world would be worse than described above if you didn't use SACK. But, if you have to take at most one loss every 100 minutes that's pretty bad already.) There is an experimental change to TCP's algorithms specified in RFC3649 (that only applies when you are going quite fast). allman -- Mark Allman -- ICIR -- http://www.icir.org/mallman/ --=-=-= Content-Type: application/pgp-signature -----BEGIN PGP SIGNATURE----- Version: GnuPG v1.2.3 (FreeBSD) iD8DBQFAT27AWyrrWs4yIs4RAu8tAJ9GguPWqUuyEUzdCoNTmECVS+dlLwCeI8zU LVuiXT0+1FHf03QjAW4NaJk= =zGI3 -----END PGP SIGNATURE----- --=-=-=--
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20040310193840.6479F77A6D4>