Dear Daniel, I tried it. Thanks for the help. It is good solution for affinity CPU core with GPU. But if I execute one gpu job at the some node, all the next gpu jobs can not use this node for calculation, although one gpu is free.
Vova. С Уважением, Владимир. 2016-04-16 0:32 GMT+10:00 Daniel Letai <[email protected]>: > This is somewhat convoluted, but you might achieve this with gres.conf > similar to > > Name=gpu CPUs=0,1 > Name=gpu CPUs=10,11 > Name=cpu CPUs=2-9,12-19 count=16 > > and when submitting a job > sbatch --gres=gpu:1 > Or > sbatch --gres=cpu:16 > Or > sbatch --gres=gpu:2,cpu:4 > > In theory the last one would consume 6 cpus and 2 gpus. > > I'm not near a cluster now, so this is pure guesswork. > > > On 04/15/2016 12:50 PM, Vladimir Goy wrote: > > Dear Developers, > > I study Slurm now. About two weeks I can not solve the next problem: I > need to separate Node (20 CPU + 2GPU) into 2 parts: Node0 -- 16 CPU, > without GPU; Node2 -- 2GPU + 4CPU (it will be good, if it will be CPU > cores: 0,1, 10,11, for best affinity with GPU). Also I need to make 2 > partition: cpu, gpu. I would like run single-process tasks in the partition > gpu/cpu. > > I use: > SchedulerType=sched/backfill > SelectType=select/cons_res > SelectTypeParameters=CR_CPU > GresTypes=gpu > > and MaxCPUsPerNode in partition declaration. Jobs for partition cpu runs > well, for partition gpu -- bad. Only 1 job per node is available. The > second jobs is pending, while one GPU is free. I do not know, how to solve > it. > > Please, can You help me? > > > Best Regards, Vova. > > >
