Hello Bjørn-Helge,
Le jeudi 8 janvier 2026 à 11:23, Bjørn-Helge Mevik via slurm-users slurm-users@lists.schedmd.com a écrit :
Rémi Palancher via slurm-users slurm-users@lists.schedmd.com writes:
Slurm does not currently provide native enforcement of consumption-based quotas (e.g. CPU or GPU minutes).
I don't think that is 100 % accurate. Slurm does have the GrpTRESMins specification, which can be set up users, accounts and QoS'es. This limits the number of cpu, gpu, memory or billing-minutes they can use.
Unfortunately, Slurm considers GrpTRESMins on associations against usage computed for fairshare, which is decayed with half life or reset period. This means this limit is released with time going on, unless you disable usage decaying completely.
One question: Does the slurm-quota tool take requeued jobs into account, so that a job that is requeued (either manually or by node failure) will not exceed the quota?
No, that's indeed one limitation. Slurm calls the completion plugin on requeued jobs, actual usage is well accounted then. However, Slurm does not recall the submission plugin on requeue (neither submit nor modify callbacks), therefore the quota can't be enforced for requeued jobs the way it's currently designed.
Best regards, -- Rémi Palancher Rackslab: Open Source Solutions for HPC Operations https://rackslab.io