Management of heterogeneous jobs in Slurm¶
Since version 17.11, Slurm has the ability to handle computations requiring heterogeneous resources. For example, a different amount of memory on certain cores or the ability to request simultaneous jobs on different partitions to perform a coupling.
Case study: coupling between 2 codes, visualization code, launching in client-server mode (slave/master)...
Launching a heterogeneous job made of 2 jobs running at the same time, on different resources, means describing these 2 jobs in the same submission script.
The submission is identical to a normal job:
Another solution, in command line:
Follow-up and stop of a job¶
The display of jobs in
squeue changes: at the slurm level, it corresponds to 1 job containing sub-jobs. So they appear under the same number with "+num" corresponding to the packjob.
To kill the job with its sub-jobs, just use
scancel on the job number (in the above example:
To kill the
+0 sub-job only, just specify
If the job is in
pending state, then it is not possible to undo the sub-jobs, only the job and its sub-jobs.
Pour plus d'information, consulter la documentation de Slurm en prenant soin de choisir la version identique à celle de Slurm sur Myria. En octobre 2022, Myria est sous Slurm 20.02.7.