Mixing Preemptible and non-preemptible partitions

[Paul, Austin]

Hi,

I have a slurm-on-gcp cluster set up 1 year ago using schedmd/slurm-gcp, and it currently has a single partition consisting of preemptible nodes with GPU attached.

I would like to add a partition consisting of non-preemptible versions of these nodes. In my slurm.conf file, I see the following lines:

# COMPUTE NODES
GresTypes=gpu
NodeName=DEFAULT Sockets=1 CoresPerSocket=1 ThreadsPerCore=2 RealMemory=7070 State=UNKNOWN Gres=gpu:1
NodeName=ajp-slurm6-compute[1-300] State=CLOUD
PartitionName=debug Nodes=ajp-slurm6-compute[1-300] Default=YES MaxTime=INFINITE State=UP LLN=yes

I'm assuming I have to add another PartitionName line, but I'm not seeing how to specify preemptibility... Any assistance would be appreciated.

(The new cluster-services scripts in the marketplace deployment look really handy, perhaps it's time I should just redeploy my cluster)

Thank you,

Austin

[Wyatt Gorman]

Hi Austin,

If you're running a 1-year-old version of the Slurm-GCP scripts you may not have partitions supported yet? You can check if there's a partition field in your YAML file. If you do, you need to modify both /apps/slurm/scripts/config.yaml and slurm.conf to add the partition you're looking for. Let me know if you can use partitions and I'll send you some instructions for adding a partition.

Otherwise you'll need to update to a newer version with partition support.

Thanks,

Wyatt Gorman HPC Solutions Manager https://cloud.google.com/hpc

--
You received this message because you are subscribed to the Google Groups "google-cloud-slurm-discuss" group.
To unsubscribe from this group and stop receiving emails from it, send an email to google-cloud-slurm-...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/google-cloud-slurm-discuss/CAE2Zhhfi0xe91Jc%2B9yd3Gxg1Q_q54oahYVneE_YYhH334Weiig%40mail.gmail.com.

[Joseph Schoonover]

Hey Austin,

Feel free to reach out directly if you need help using the marketplace deployment with the cluster-services scripts to get your system set up as needed.