On Thu, Nov 14, 2019 at 06:00:27AM +0000, Manuel Sopena Ballesteros wrote: > Dear SGE community, > > > > Is there a way to limit the number of cores used per job like h_vcpu type > of flag? > > If yes, then what would SGE do if let say a user runs something like qsub > –b myscript.sh –l h_vmem=1G h_vcpu=2 and the job tries to use 10 cores? > > >
Generally the way smp is set up under grid engine is to create a PE named smp with an allocation rule of $pe_slots and use that wehn submitting smp jobs. Modern versions of grid engine (SoGE 8.1.9,Univa Grid Engine) have support for cpusets on Linux which can hard limit the job to using the allocated cores although you have to look out for interaction with the -binding option to qsub. With that said we don't use the SoGE builtin cpuset support at UCL in part because some of our clusters still run SoGE 8.1.8 where cpuset support is buggy and in part because we also use cgroups to control memory usage and prefer a unified approach. We use a prolog script that sets up the cpuset (and memory cgroup) and chgrps it to the unix group assigned to the job by SoGE This lets the unprivileged starter method drop itself into the cgroup/cpuset which is inherited by the job. If you're not running linux or are running an even more retro version of SGE then you could try qsub -binding or setting up queues that have the processors entry configured. However the binding provided by these options can be undone by the job AIUI (unlike cpusets/cgroups). William
signature.asc
Description: PGP signature
_______________________________________________ users mailing list users@gridengine.org https://gridengine.org/mailman/listinfo/users