Hi,

Am 29.04.2013 um 10:00 schrieb [email protected]:

> we have parallel jobs with mixture openmp and mpi compiled with ifort. We 
> found out that job threads change the physical processor core from time to 
> time, this can lead to increased run time of the job. Outside the gridengine 
> there is a variable "export KMP_AFFINITY=scatter " which set affinity to 
> physical cores of the threads.
> Unfortunately we can not use this setting in our gridengine job scripts. With 
> this definition two different jobs which are submitted to the same node will 
> occupy the same processor cores
> by the gridengine.
> Does anyone know a way to define this option in the gridengine directly ? I'm 
> using ogs 2011.11p1 on Intel 6core processors.

With 6 cores per machine I would simply request full nodes.

AFAIK the commercial GridEngine from Univa contains extensions to monitor 
granted cores on a host level, so that each of such a job could be given an 
unique list of cores. If any KMP_... variable would allow a syntax to accept a 
list of cores, it should work that your executable can make a proper 
allocation. But such a thing is not available in the open source version.

What might work: define one queue per core with only one slot. Hence you will 
get many different queues per node for a parallel job and from the $PE_HOSTFILE 
you know which cores can be used where.

-- Reuti
_______________________________________________
users mailing list
[email protected]
https://gridengine.org/mailman/listinfo/users

Reply via email to