bob,
i have also few comments. i wasn't talking about cputime - walltime
differences. i usually inspect such things using the xpbs frontend to
the pbspro queuing system on clusters i use. because i use mpiexec to
"propagate" the job processes among the nodes, pbspro tells me not only
cputime and walltime (cputime should be walltime * number of
processors), but also cpupercent usage. in the ideal case, it should be
number of nodes * 100.
what i am seeing for one of my systems (btw, it is the system for which
i was "complaining" for getting only 10ns per day some time ago :) is
that the cpupercent is 140-160 only. usually this is a sign that there
is probably a dead job eating cputime at some of the allocated nodes,
but in this case, all processors are free... so since there is heavy
communication for my system (it is very small, but i need very long
simulation times), i thought myrinet could help... and it helps, i have
cpupercent at 197 for 2 nodes = almost perfect situation...
regards,
--
Lubos
_._"
-----------------------------------------------------------------------
The AMBER Mail Reflector
To post, send mail to amber.scripps.edu
To unsubscribe, send "unsubscribe amber" to majordomo.scripps.edu
Received on Fri Apr 23 2004 - 19:53:01 PDT
