[slurm-users] issue with --cpus-per-task=1
Diego Zuccato
diego.zuccato at unibo.it
Thu Mar 10 11:56:57 UTC 2022
Hi Benjamin.
IIUC, what you're seeing is due to multithreading. If you ask for a cpu,
you get allocated a core, that means it can handle two tasks. Hence the
"payload" is run twice to fill the two tasks. If you also specify you
need only one task, it gets run only once (but should still be able to
multithread and use the second thread on the same core).
Diego
Il 10/03/2022 12:23, Benjamin Glaessle ha scritto:
> Hello all!
>
> we are using slurm 20.11.8 with
>> SelectType = select/cons_tres
>> SelectTypeParameters = CR_CORE_MEMORY
> and nodes with enable hyperthreading, e.g.
>> NodeName=slurm-node?? NodeAddr=192.?? Procs=72 Sockets=2
>> CoresPerSocket=18 ThreadsPerCore=2 RealMemory=...
>
> when launching jobs on these nodes with --cpus-per-task 1 they execute
> twice:
> > $ srun --cpus-per-task 1 echo foo
> > foo
> > foo
>
> digging deeper I found
>> $ srun --cpus-per-task 1 env | grep -i tasks
>> SLURM_NTASKS=2
>> SLURM_TASKS_PER_NODE=2
>> SLURM_STEP_NUM_TASKS=2
>> SLURM_STEP_TASKS_PER_NODE=2
>> SLURM_NTASKS=2
>> SLURM_TASKS_PER_NODE=2
>> SLURM_STEP_NUM_TASKS=2
>> SLURM_STEP_TASKS_PER_NODE=2
>
> whereas `scontrol show job 12345 | grep -i -e numtasks -e numcpus` for
> both "env" and "echo" job gives
>> NumNodes=1 NumCPUs=2 NumTasks=1 CPUs/Task=1 ReqB:S:C:T=0:0:*:*
>
> A test node without ThreadsPerCore=2 behaves "normally".
> Also
> > $ srun -n1 --cpus-per-task 1 echo foo
> > foo
> resolves the problem.
>
> This seems like a bug to me.
> Can this be reproduced (on newer versions)?
>
> Can this somehow be avoided by setting a default number of tasks or some
> other (partition) parameter? Sorry for asking but I couldn't find
> anything in the documentation.
>
> Let me know if you need more information.
>
> Best Regards, Benjamin
>
--
Diego Zuccato
DIFA - Dip. di Fisica e Astronomia
Servizi Informatici
Alma Mater Studiorum - Università di Bologna
V.le Berti-Pichat 6/2 - 40127 Bologna - Italy
tel.: +39 051 20 95786
More information about the slurm-users
mailing list