Messages in this thread | | | Date | Mon, 24 May 1999 16:34:17 -0400 | From | Josip Loncaric <> | Subject | Re: [patch] TCP/IP delacks disabled with MPI |
| |
"David S. Miller" wrote: > > Date: Fri, 21 May 1999 19:56:29 +0200 (CEST) > From: Andrea Arcangeli <andrea@suse.de> > > Right now the low bound is the RTO that is lowbound to 200msec > itself. If you decrease the lowbound of the rto then also the ato > will decrease. > > And unless you come up with some clever idea, if you implement > decreasing the RTO lower bound you will crap out completely with all > BSD'like TCP stacks where this is the smallest resolution of their TCP > timers. Read the commentary in tcp_input.c about "HZ/5" etc. > > I cannot allow this to change in the standard kernel unless a > resolution to this issue is discovered.
This should not be changed in the *standard* kernel but should be available as a socket option so that people who build clusters using switched Fast Ethernet can get reasonably responsive TCP. We do this by changing the default "HZ/5" to the more responsive "HZ/50" whenever TCP_NODELAY socket option is set.
BTW, there are other timeout values which can and should be changed for clusters using TCP_NODELAY sockets (e.g. the full second retransmit timer when sh->users is set and another seldom used 30 second partial packet timer).
These issues are of crucial importance for Linux based clusters, which use TCP in ways very different from "standard". Improving a few timeout constants is a start, but in the future, we'll probably see more elaborate TCP enhancements driven by the needs of parallel processing applications.
Sincerely, Josip
-- Dr. Josip Loncaric, Senior Staff Scientist mailto:josip@icase.edu ICASE, Mail Stop 132C http://www.icase.edu/~josip/ NASA Langley Research Center mailto:j.loncaric@larc.nasa.gov Hampton, VA 23681-2199, USA Tel. +1 757 864-2192 Fax +1 757 864-6134
- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.rutgers.edu Please read the FAQ at http://www.tux.org/lkml/
| |