AnsweredAssumed Answered

how to avoid performance loss on AMD cluster by strictly fillingupstopping other jobs from entering the nodes whereupon I am running (Sun Grid Engine 6.2u5p2/openMPI 1.4.4) :: in the past, submission under "qsub -l exclusive .." was possible; not anymore

Question asked by mperezjigato on May 24, 2016



I am running my application on an AMD cluster containing 32 nodes with 16 core each. I am using SGE 6.2u5p2 and openMPI 1.4.4, and noticed that all my runs suffer a huge loss of performance (even 40 percent) whenever other jobs run on my running nodes (typical run uses 48 cores and is intended for 3 nodes each with 16 cores; it is never the case..)


Is there any strict procedure as to avoid the performance loss? I guess, the best would be filling up all 16 cores per node systematically. The "-l exclusive" option inside my submission script (or else in qsub) does not exist in my current version


thank you