We noticed an inconsistency when slurm has pending jobs in the number of expected nodes it will use. My understanding is that the number should be calculated based on the maximum node size (by cpu) and the number of CPUs a job needs.
A quick and dirty review of the code makes it look like it min_nodes is calculated from num_tasks, and does not take into account the # of CPUs per task. I could be missing something, however.
Many of our jobs are hybrid MPI + OpenMP. For example, the jobs might use 4, 5 or even 10 CPUs per task. This causes the minimum node calculation to seemingly be off by that factor.
Best, Jesse Stroik
