Path: utzoo!attcan!uunet!lll-winken!decwrl!ucbvax!MANTA.NOSC.MIL!ron
From: ron@MANTA.NOSC.MIL (Ron Broersma)
Newsgroups: comp.protocols.tcp-ip
Subject: Re: *.jhuapl.edu -- serious gateway thrashing
Message-ID: <9003020608.AA10396@manta.nosc.mil>
Date: 2 Mar 90 06:08:31 GMT
Sender: daemon@ucbvax.BERKELEY.EDU
Organization: The Internet
Lines: 33

I'm wondering if some of this gateway thrashing isn't related to the
fact that EGP packets from the MILNET core started exceeding 4096 bytes
a month or two ago.  At that time, I was tracing some thrashing problems
and I noticed the following symptoms.  Over the course of an hour, the
packets would gradually increase in size.  Just as they got within 10
to 20 bytes of 4096, many of the EGP implementations would suddenly
start getting checksum errors or buffer overflows because they had
4K buffers.  The ones that got a few packets with bad checksums would
suddenly stop peering with that core gateway.  Then, all of a sudden the
EGP packets out of the core would be smaller by a few hundred bytes because
of many fewer peers.  As the EGP players all tried to acquire a different
gateway, they would not get the checksum errors for an hour or so until
the packets approached 4096 bytes and they would again perform this
dance-of-the-gateways.

The message here is to make sure your EGP implementation can handle
packets larger than 4K bytes.  The most recent gated supports 8K packets
as I recall.

Something I had considered was to make a list of all the networks that
disappeared from the EGP packets right after the "dance".   Then if one
could determine who announced those nets to the core you could get a
handle on where the broken EGP implementations were located.

There's some other strangeness going on too.  We had a case this week
where one site running EGP was announcing its network to the core but
the core wasn't telling anybody else about it.  By peering with a
different mailbridge, it started working.  Strange.

And to top it off, the ground started shaking yesterday.  But we think
that is an unrelated (hardware) problem.

--Ron