On Fri, Feb 09, 2007 at 08:03:54AM +0000, Michael Rogers wrote: > Matthew Toseland wrote: > > Fast retransmission = If I get acks for 2,3,4 but not 1, then I resend > > 1, after X RTTs have elapsed? > > TCP fast retransmits 1 when 2,3,4 have been acked, without considering > the RTT. The simulator code also checks that 1 was last sent > 1.5*RTT+MAX_DELAY seconds ago, but this is just to prevent 1 from being > retransmitted again when 5 is acked.
Okay so TCP has a cooldown instead? Won't resend a given packet more often than once per RTT? > > > Why would it be more than 4*RTT+MAX_DELAY? > > Think of a probability distribution with mean m - some part of the tail > extends beyond 4*m. The question is, does that part of the tail contain > more than 5% of the samples? It depends on the variance. Indeed. > > We could change it to 10*RTT+MAX_DELAY, for example, and have a pretty > good chance of avoiding unnecessary retransmissions - but the packets > that were transmitted would arrive very late. > > Cheers, > Michael -------------- next part -------------- A non-text attachment was scrubbed... Name: signature.asc Type: application/pgp-signature Size: 189 bytes Desc: Digital signature URL: <https://emu.freenetproject.org/pipermail/devl/attachments/20070209/49c9e96c/attachment.pgp>
