There was a thread a short while ago about some TCP performance issues
that someone was having during a little throughput test they were
doing. They were using a tool called "traffique" to get their
benchmarking done.

Someone mentioned that there were some checksum code problems that
were only recently fixed--

A while back i presented to the tech-net@ list a set of tcpdumps from
two machines and was comparing tcp recovery speeds during packet loss
(which NetBSD seemed to be having some trouble with). These recovery
speeds (if that's what they were) were affecting throughput on bad
network lines significantly.

It was suggested at one point that Linux (the one without this
trouble) was breaking RFC-defined behaviour.

(Reference the notes here:
http://mail-index.netbsd.org/tech-net/2000/09/ )

Would these two possibly be related? My problem was *purely* in the
delays it took before retransmitting (palpably about a half second to
.8 seconds.)

Marc