If you are running SGE 6.2u5, then you can use exclusive job scheduling: http://docs.oracle.com/cd/E24901_01/doc.62/e21978/management.htm#sthref431
IMO, it is cleaner than other ways. Also, some OpenMP implementations assume that they get the whole node, and thus it is better to just use the node exclusively. Rayson On Fri, May 11, 2012 at 11:33 AM, Jack Stalnaker <jack.stalna...@gmail.com> wrote: > Hi, > > I am trying to determine the best way to ensure that an SMP (OpenMP) job > runs on a single node of the cluster exclusively. That is to say I want the > entire node to this one job, even if the job doesn't use all cores on the > node. Our cluster has 12 physical cores per node. My predecessor achieved > this by setting the parallel environment as -pe smp 12, and setting the > OMP_NUM_THREADS variable to whatever number of threads he wanted to actually > execute on in the submitted script. This seemed to work well until recently. > A job submitted in this way went onto a node where another job was already > running, though it doesn't seem it should have, since there shouldn't have > been 12 slots available on that node. > > Can anyone tell me why this might've happened? Curiously, the cluster was > tapped out at this point, and quse showed that the number of free slots went > to -12! I had never seen that before either. > > I'm not sure if it's related, but qmon shows 24 CPUs per node, rather than > 12 under queue control->queue instances. I'm guessing that is due to hyper > threading. > > Thanks. > > _______________________________________________ > users mailing list > users@gridengine.org > https://gridengine.org/mailman/listinfo/users > _______________________________________________ users mailing list users@gridengine.org https://gridengine.org/mailman/listinfo/users
