[Bi-users] Issue with Slurm task allocation after yesterday's update (now fixed)
Torbjörn Lönnemark
ketl at nsc.liu.se
Thu Dec 10 21:23:41 CET 2020
Dear Bi users,
after the upgrade yesterday, a change in behaviour in the Slurm
job scheduler used on Bi caused jobs to be allocated with twice the
number of expected tasks per node - one per hyperthread instead of one
per core.
This affected jobs where *only* a node count (-N/--nodes) was specified
when they were submitted (but no -n/--ntasks-per-node/etc).
Affected jobs may have suffered from lower than expected performance,
or out-of-memory errors.
We have now updated Slurm to restore the previous behaviour, so
requesting only a node count when submitting a job should now work
as expected again.
Best regards,
Torbjörn Lönnemark
National Supercomputer Centre (NSC)
More information about the Bi-users
mailing list