loki@NAZGUL.PHYSICS.MCGILL.CA (Loki Jorgenson Rm421) (04/25/91)
OK... no one gave me any particularly useful info about what to
look for insofar as our net problem goes (recap: high collision rates
on 4D IRIS servers (2-10%) compared to SUNS (0.1% on servers and clients)
according to "netstat -i"). Examining net statistics shows that the
bulk of the problem lies with NFS to/from IRI. SUN servers have no such
problem regardless of whether they are serving IRI or SUN.
Besides the usual thinnet cabling limitations (max 185m of cable
segment before repeater is necessary, 50cm multiples in length, max 30
machines), the only option suggested was to have a Sniffer or NetVisualizer
available, neither of which is financially feasible for us. No one
mentioned a minimum cable length (other than 50cm).
However.... at least two other sites noted the same problem and
have failed to come up with an explanation. The Geometry hotline has yielded
nothing. I was wondering if someone at SGI would like to comment.
Or at least propose a good reason why this should be acceptable from an IRIS.
???
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) 398-7027vjs@rhyolite.wpd.sgi.com (Vernon Schryver) (04/26/91)
In article <9104250127.AA28210@nazgul.physics.mcgill.ca>, loki@NAZGUL.PHYSICS.MCGILL.CA (Loki Jorgenson Rm421) writes: > > OK... no one gave me any particularly useful info about what to > look for insofar as our net problem goes (recap: high collision rates > on 4D IRIS servers (2-10%) compared to SUNS (0.1% on servers and clients) > according to "netstat -i"). 3.3 and all previous versions of IRIX count both "deferals" and "collisions" in the Coll column in `netstat -i`. A deferal happens when a machine decides to transmit and discovers that the media is already busy. A collsion happens when a transmitter notices that some other station is transmitting while it is also transmitting. Only collisions are counted in the Coll column in 4.0. This ancient characteristic (I'm reluctant to call it a "bug" not just because I did it but because the deferal rate seems useful) was noticed during testing of new models that can transmit arbitrarily long strings of packets with 9.6 usec gaps. They drove the "Coll" rate up to 90% on other machines. 10% is not a high collision rate, even if you don't count deferals. Vernon Schryver, vjs@sgi.com