Skip to content

Updated midas jobscript for freja (fixes #95)

Joakim Löw requested to merge joakim.low/midas:fix-jobscript-freja into master

I've modified the "example" jobscript to fit freja:

  • Fewer worker nodes
  • Updated comments for freja
  • Changed the MEM_PER_WORKER calculation. This calculation must have been faulty for bi, for any value on cpus-per-task other than 2, but should now distribute the memory properly across workers.
  • New Mambaforge module version
  • New dask command form for scheduler and worker (previous form deprecated)
  • I also had to change the reference period in the example, since the example data does not seem to contain the full period range.

The script still uses exclusive nodes for scheduler and client. Sharing nodes for schedulers seems to be a still open issue for freja (also not exclusive to dask). I did try a quick test sharing node between scheduler and client, but failed, and can't put more time into this now, so I think this example is as good as we can do for the moment.

Merge request reports