Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!purdue!tut.cis.ohio-state.edu!ucbvax!DAINICHI.TN.CORNELL.EDU!swb
From: swb@DAINICHI.TN.CORNELL.EDU (Scott Brim)
Newsgroups: comp.protocols.tcp-ip
Subject: Re: routing protocols
Message-ID: <8911201745.AA25525@chumley.TN.Cornell.EDU>
Date: 20 Nov 89 17:45:46 GMT
References: <42691@bu-cs.BU.EDU>
Sender: daemon@ucbvax.BERKELEY.EDU
Organization: The Internet
Lines: 17


   >Oh, I don't know.  The first NSFnet backbone risked a delay/congestion
   >based algorithm on an operational (but noncommercial) network.

The implementation of the HELLO protocol used in the first NSFNET
backbone (on fuzzballs, by Dave Mills) was supposed to be independent
of load on the net, in that the routing messages were generated just
barely outside of the interface cards themselves.  There turned out to
be some wobble, and Dave put a lot of work into controlling and
damping the incredible thrashing that occasionally came about due to
that wobble (also gated had a sliding window and filter put in to damp
it even further).  I don't know anyone who's working on dealing with
these problems for large nets anymore.  The main problem is that
you're in a situation where the system can (and probably will) change
direction faster than you can detect a wobble, send out an adjustment
to it, and have your adjustment stabilize.  You just keep thrashing.
							Scott