Martin,
it's an allocation issue. With CR_CPU I can run 32 different tasks on a
16 core system with hyperthreading. With CR_Core I can run 16 tasks
since each task allocates 2 CPUs instead of 1. This particular
application does not care if it shares a core but it does care about how
many single task jobs are able to run simulatiously. In fact I do want
to run multiple jobs on a single core, with single job per cpu.
So I can't use the CR_Core_Memory since I need the full 32 cpus
allocated to 32 different tasks and I can't use the CR_CPU_Memory not
supported in the partition.conf.
Any other option?
Thanks
Eva
On Fri, 9 Aug 2013 [email protected] wrote:
> Eva,
> Does it matter? CR_Core and CR_Core_Memory prevent multiple jobs from
> being allocated CPUs on the same core. So whether Slurm allocates one or
> both CPUs on a core to your job shouldn't make any difference for other
> jobs. Or is this just an accounting issue?
> Martin
>
>
>
> From: Eva Hocks <[email protected]>
> To: "slurm-dev" <[email protected]>,
> Date: 08/09/2013 11:41 AM
> Subject: [slurm-dev] CR_Core_Memory and CR_Core versus
> CR_CPU_Memory
>
>
>
>
>
>
>
> I am struggling to set slurm to allocate the correct CPUs for a job
> requesting 1 task.
>
> With hyperthreading enabled and the CR_Core setting slurm allocates 2
> CPUs per job requesting 1 CPU:
>
>
>
> srun -n 1 --pty --ntasks-per-node=1
>
> shows:
> NodeList=gpu-1-9
> NumNodes=2 NumCPUs=2 CPUs/Task=1
>
>
>
> The node configuratoin has the correct settings:
> CPUs=32 Sockets=2 CoresPerSocket=8 ThreadsPerCore=2
>
>
>
> I can't use the CR_CPU_Memory since I would like 2 partitions to monitor
> CR_Core_Memory and one partition to monitor CR_Core. The partition
> SelectTypeParameters does not allow for CR_CPU
>
>
> How do I configure slurm to allocate the correct CPUs requested by the
> job?
>
>
> Thanks
> Eva
>
>
>
