Closed xylar closed 8 months ago
With this fix, Polaris tasks that use more than one Frontier node run successfully, whereas they fail because they try to run on 64 cores per node (the total, rather than the allocatable number) without this fix.
This needs to be tested on Chrysalis and Perlmutter to make sure it doesn't break anything there.
This approach didn't work on Compy.
Some systems like Frontier have cores that aren't allocatable. These need to be excluded from the core count that Polaris determines from slurm.
Checklist
Testing
comment in the PR documents testing used to verify the changes