Separate partition for ondemand jobs?

General Discussion
1

Hello, looking for some on advice on whether we should mix ondemand jobs, which for us are predominately GUI/interactive type of workloads, with batch jobs in the same slurm partition(s) or to have separate partitions based on workload type?

2

This is what we have in place:

PartitionName=ondemand  Nodes=ucs,amd,gpu   Priority=10 MaxTime=2-0 TRESBillingWeights="CPU=1.0,Mem=0.084G,GRES/gpu=19.695" Default=NO MaxNodes=1 Hidden=YES
PartitionName=login     Nodes=login         Priority=30 MaxTime=2-0 TRESBillingWeights="CPU=1.0,Mem=0.084G"                 Default=NO MaxNodes=1 Hidden=YES

which limits the OOD jobs running on the ondemand partition to a maximum of 2 days. Interactive apps like Matlab that may have longer running jobs are permitted to run on our regular partitions.

The login partition has just about the highest priority, and allocates jobs on the actual login nodes. This is the last resort for when the cluster is super busy, we can still let users allocate Code Server or some other popular GUIs that require immediacy.

For our classroom offerings, we have the “Education Cluster” which is just a reservation:

ReservationName=EducationCluster StartTime=2025-08-25T00:00:00 EndTime=2025-12-13T00:00:00 Duration=110-01:00:00
   Nodes=cpu-intel-[63-78] NodeCnt=16 CoreCnt=704 Features=(null) PartitionName=(null) Flags=IGNORE_JOBS,SPEC_NODES,MAGNETIC
   TRES=cpu=704
   Users=(null) Groups=(null) Accounts=classes Licenses=(null) State=INACTIVE BurstBuffer=(null) Watts=n/a
   MaxStartDelay=(null)

This reservation is put into place for the duration of the semester and is magnetic for the Slurm account classes. We make a Slurm account for each class, like phy266, that is also a sub-account of classes. Any job that is submitted under any of those class accounts automatically gets pulled into the Education Cluster, and are instantly allocated.

1 Like
3

Hi and welcome!

Sorry for the delay. It’s up to you really. We (OSC) don’t have special partitions but others may.

For what it’s worth: Most interactive jobs at OSC allocate only 1 core. I would imagine Slurm is better at bin packing than any human. Which is to say, it’s likely to utilize your resources when they’re in a shared pool. The one thing I’d worry about in an interactive specific partition is under utilization.

4

Thanks very much all for replies!