loki@NAZGUL.PHYSICS.MCGILL.CA (Loki Jorgenson) (06/09/91)
Hey ho....
This summary is really late (2 months) but I thought it might be
useful to someone else someday.
Originally, I had observed collisions on our SGI machines
(4D/20, 4D/25 running 3.3.1) to be about 4-10% of the number of output
packets (according to "netstat -i"). This relative to <1% on our SUN3s
on the same net. Of the many proposed origins of the problem, here are
the ones that seemed the most relevant:
1) Recent changes in the local network topography had created problems
related to over-long net segment, bad connections or other technical bugs.
2) Unusual hardware arrangements: cascaded multiport transceivers or
the like.
3) SQE heartbeat on the transceivers was accidentally on (it shouldn't
be according to SGI).
4) SGI claims that their machines are more "sensitive" to errors in the
sense that events are counted as errors which many other brands of
machine ignore. This doesn't affect anything except how netstat keeps
statistics.
After several months of investigation, our local problems were
three-fold. One, we didn't know about 4); this was the major non-problem.
Two, many of our SGI transceivers were indeed configured with SQE on.
Three, our local net segment length was becoming critically long; we
have since installed a repeater to improve the overall net performance.
My thanks to (hope I didn't forget anyone):
Reinhard Doelz / Biocomputing Basel <doelz@yogi.vmsmail.unibas.ch>
Paul Hilchey <hilchey@ucs.ubc.ca>
hogg@ucs.ubc.ca (John Hogg)
amys@ncsa.uiuc.edu (Amy Swanson)
Bernard Daoust (local sales guy)
Regards,
_ _ _ _
Loki Jorgenson / / _ _ _ _ _ \ \ node: loki@Physics.McGill.CA
Grad/Systems Manager /_/_/_/_/ \_\_\_\_\ BITNET: PY29@MCGILLA
Physics, McGill University \ \ \_\_\_/_/_/ / / fax: (514) 398-3733
Montreal Quebec CANADA \_\_ _/_/ phone: (514) 398-7027
-* Anatomically correct *-