Reuti wrote:
Am 11.04.2012 um 09:52 schrieb Ursula Winkler:
Ron Chen wrote:
As long as Grid Engine is not using more job slots than the configured number
(the no. of processors in the system by default), then it is the correct
behaviour.
The OS should spread the tasks around evenly. How many tasks are there on your
12-core machine?
We have on each host 2 6-core Prozessors, ergo 12 slots. And 12 slots are
configured in the SGE for each host. The SGE also correctly does not allow more
than 12 tasks, but this 12 tasks are not distributed in the way that each one
gets a separate core, but e.g. 8 get the cores 0-3,
These go to one MPI job or to two?
two mpi-jobs (one with 4 the other with 8 tasks): the first get the
cores 0-3, the second get 0-7 and not 4-12
4 tasks get cores 4-7 and the other cores remain unused. The double seizure
always occurs when more than 1 MPI-Job (not single prozessor jobs) are
scheduled to the same host.
This could also be a problem of the MPI implementations. Which one do you use -
you use a plain mpiexec?
Yes (mpiexec by mavpich2 1.8)
Ursula
_______________________________________________
users mailing list
[email protected]
https://gridengine.org/mailman/listinfo/users
