[2931] in linux-net channel archive
More on slow TCP in the pre-2.0 kernels
daemon@ATHENA.MIT.EDU (Eric Schenk)
Fri May 17 07:47:31 1996
To: linux-net@vger.rutgers.edu
Date: Fri, 17 May 1996 02:43:11 -0400
From: Eric Schenk <schenk@rnode84.cs.toronto.edu>
I mentioned in another post that the current TCP networking code will
not perform as well as the 1.2.13 code under some pathalogical conditions.
In particular, now that we have fast retransmits and various corrections
to the Jacobson slow start and congestion avoidance code, it turns out
that if we are transmitting over highly congested links our performance
goes down from 1.2.13 levels.
The problem is that if there is never any gain from growing the congestion
window from size 1 to size 2, we will spend a lot of time oscilating the
congestion window between 1 and 2, and when fast retransmit gets triggered
we will inject a long sequence of packets into the network which will all
be lost. This just generally results in very slow behavior under these
circumstances. At the moment I cannot see how we can avoid this without
diverging from Jacobson's algorithms in some signficant way.
An alternative is to provide a compile time option to turn the fast
retransmit code off. This may keep some people who usually work over
highly congested links happier. Can anyone suggest anything else
we can do to help with this?
-- eric
P.S. I won't see any suggestions until next Tuesday or Wednesday,
so you'll just have to discuss the merits of suggestions amongst
yourselves till them :-)
---------------------------------------------------------------------------
Eric Schenk www: http://www.cs.toronto.edu/~schenk
Department of Computer Science email: schenk@cs.toronto.edu
University of Toronto
