RE: Connected mode draft and retransmit implosion
Ajoy Siddabathuni <[email protected]> Fri, 13 Aug 2004 18:04:42 -0700 (PDT)
| Newsgroups | gmane.ietf.ipoib |
|---|---|
| Message-ID | <Pine.GSO.4.05.10408131759380.3910-100000@endeavour> |
Isnt tcp timers in the order of msec and ib timers in the order of usecs? I think its just a problem of tuning the stack for the application in question. ~Ajoy On Fri, 13 Aug 2004, bill wrote: > Yes, that is exactly the problem... Both layers retransmitting at the > same time > > And possibly not recovering for quite a while > Bill > > > -----Original Message----- > > From: [email protected] > > [mailto:[email protected]] On Behalf Of Fab Tillier > > Sent: Friday, August 13, 2004 5:42 PM > > To: 'bill'; [email protected] > > Subject: RE: [Ipoverib] Connected mode draft and retransmit implosion > > > > > > > From: bill [mailto:[email protected]] > > > Sent: Friday, August 13, 2004 4:57 PM > > > > > > Vivek, > > > > > > So I have been concerned about the Connected mode draft and > > I run into > > > a simple problem here... > > > > > > Imagine I have a TCP socket running over a 64K RC MTU link. > > > > > > I transmit the IP packet (all 64K of it) > > > > > > Infiniband breaks the packet up into a bunch of 2K Infiniband MTU > > > packets (all 32 of them) - lets say I loose the second packet. > > > > > > No problem IB will retransmit the packet over the link and > > it gets to > > > the other side... > > > > > > The problem that I want to see - is what happens if the TCP > > retransmit > > > timer goes off - causing the whole TCP segment to be retransmitted. > > > > Wouldn't you end up with the same packet getting sent twice? > > If so, why is this problematic (aside from the inefficiency)? > > > > - Fab > > > > > > > > _______________________________________________ > > IPoverIB mailing list > > [email protected] https://www1.ietf.org/mailman/listinfo/ipoverib > > > > > > _______________________________________________ > IPoverIB mailing list > [email protected] > https://www1.ietf.org/mailman/listinfo/ipoverib >