On Wed, Mar 10, 2004 at 06:23:07PM +0200, Petri Helenius wrote: > Brooks Davis wrote: > > >For that matter, there are sufficent drops on 10GbE from data errors to > >insure that two boxes connected back to back won't achieve line speed on > >a single TCP session. > > > Do you have data to back this up or are you just using broadcom chipsets?
I looked for the paper I paraphrased, I'm pretty sure if was one by Sally Floyd. I didn't find it, but this paper talks a bit about the issue on page 7: http://www.calit2.net/research/labs/features/CACM/CACMDefanti.pdf The problem is that the BER of a typical optical link is high enough that the link will almost certantly discard at least one packet before you get out of slow-start and once that happens it, AIMK means it take hours or even days to get back up to the top even assuming you don't lose further packets. This isn't a problem for most people, but it's definalty a problem for the HPC community. -- Brooks -- Any statement of the form "X is the one, true Y" is FALSE. PGP fingerprint 655D 519C 26A7 82E7 2529 9BF0 5D8E 8BE9 F238 1AD4
pgp00000.pgp
Description: PGP signature