From: Tomek Jarosinski; Tuesday, March 16, 2004 9:32 AM
>
> I am configuring one Linux Cluster - 16xP4HT 2.8 + 1x2Xeon HT 2.8 with
> Suse 8.2 and Lam 656. Lam testsuite is running OK. When i am testing
> some lam c programs, then the cpu usage an all nodes is about 60% and
> when i am doing some LSDyna tests, than it is much worse - cpu usage
is
> only 3-7%. The same programm is running faster on one host ...
You should be able to run LSDyna at fairly high rate. BTW, are you
running SP or DP?
> All nodes have Intel Boards D875PBZ with e1000 onboard are connected
I have same mobo w/ a 2.8 P4c at home -- a very speed system indeed.
I'm running RH9.
> with GB Eth using a very good Cisco 24xGB Eth . I am thinking that
this
> low cpu usage is caused by low network performance, but i was already
> playing with different MTU, RxIntDelay,TxIntDelay,wmem_max and
rmem_max,
> there is no significant change, and average answer time between nodes
> (ping -f -s 64000 host) is ~1.5ms, so it about 80 MB/sec, probably the
> possible maximum.
Have you looked at the network media?
Do you have another switch to try out?
Alternatively, could you plug two systems together via a crossover and
try out your ping again and then maybe a two-node test?
--
David N. Lombard
My comments represent my opinions, not those of Intel Corporation.
|
