[3905] in linux-net channel archive
Re: FYI; Expotential TCP Backoff (fwd)
daemon@ATHENA.MIT.EDU (Rob Janssen reading Linux mailingl)
Tue Jul 30 08:37:56 1996
From: linux@pe1chl.ampr.org (Rob Janssen reading Linux mailinglist)
To: schenk@cs.toronto.edu (Eric Schenk)
Date: Tue, 30 Jul 1996 09:52:27 +0200 (MET DST)
Cc: ecki@lina.inka.de, linux-net@vger.rutgers.edu
In-Reply-To: <96Jul29.191255edt.15388@dvp.cs.toronto.edu> from "Eric Schenk" at Jul 29, 96 07:12:51 pm
Reply-To: linux-vger@wab-tis.rabobank.nl
According to Eric Schenk:
> I've got a copy of this. I've given it a cursory reading.
> My first impression is that there may be a good idea underneath it,
> but the mathematical argument seems very weak to me. My guess
> is that if this proposal were implemented as is on a wide scale we would
> see massive congestion collapse problems. Basically the paper
> is claiming that Van Jacobson's analysis is wrong, and that expoential
> backoff is not necessary. One or the other paper must be in error.
> At this point I'm not betting against Jacobson, although I do plan
> to give this issue a serious look as an analysis problem.
> Heck, it may be worth a paper or two :-)
Everyday practice with amateur packet radio using "TCP linear backoff"
(implemented in the JNOS package) shows that this is the wrong thing to
do. It causes congestion collapse, and because of the buggy implementation
in JNOS it even causes the connections to fail unnecessarily...
Please *don't* put something like that in Linux!
OTOH, what would be nice to have in Linux is some way of triggering the
timer when something went wrong and you know it has been rectified.
(known as "kick" in the KA9Q and derived packages)
In Linux networking you just have to wait until it feels like retransmitting
and starting the information flow again.
Rob
--
+------------------------------------+--------------------------------------+
| Rob Janssen pe1chl@amsat.org | BBS: +31-302870036 (2300-0730 local) |
| AMPRnet: rob@pe1chl.ampr.org | AX.25 BBS: PE1CHL@PI8WNO.#UTR.NLD.EU |
+------------------------------------+--------------------------------------+
