Am 22.08.2012 um 22:50 schrieb Henrichs, Juryk: > Hi, > >> <compose-unknown-contact.jpg> Reuti Mittwoch, 22. August 2012 22:19 >> Hi, >> >> Am 22.08.2012 um 21:39 schrieb Henrichs, Juryk: >> >> >>> we have a heterogeneous cluster consisting of nodes with 32 and 48 >>> cpu's. Some of our jobs are best run either on one complete node or >>> distributed over a number of nodes. In either case it is desired that >>> the job runs exclusively on each node and uses all available slots. >>> >>> The reason for this is that we use mpirun with --mca mpi_paffinity_alone >>> 1 as this gives the best performance. >>> >>> The problem is, that up to now we have to take care on which type of >>> nodes a job is running and that jobs are not distributed over mixed nodes. >>> >> >> You mean only 32 core machine or 48 core machines? >> > Yes.
Then a fixed allocation rule could do it. >>> Is there a way to submit jobs with a somehow flexible number of slots >>> such that only complete nodes are allocated? >>> >> >> Sure, you can specify a slot range for the PE request in `qsub`. >> > How is determined on how many slots the job is finally started? Available > slots depending on the requested resources? Can I be sure to end up with > completely filled nodes? Yes. And as the PEs offer only a multiple of 16 resp. 48, you can't get anything else. Submitting a job requesting something like 33 cores would never start. -- Reuti _______________________________________________ users mailing list users@gridengine.org https://gridengine.org/mailman/listinfo/users
