We have heterogeneous partitions too.  We see this occasionally, but it’s not a huge problem.  The way we have things set up is all the nodes are shared by three partitions; short-gpu, medium-gpu and long-gpu.  The difference between the partitions is the priority and the partition QoS.  Short-gpu has the highest priority, and allows the highest proportion of the GPUs to be used by a single user, but has short maximum time limit for the jobs (2 hours).  Conversely, long-gpu doesn’t let the user use many GPUs, but they can run for a long time.  Medium-gpu, obviously, is somewhere between the two.    This seems to work reasonably well, and I can usually get a GPU for a short job almost immediately.

I would check your priority weights - if you have job age dominating in the priority calculation, you’re likely to have issues where young jobs don’t run, even if they fit, with the resulting situation being what you see.  We try to set priority so that Fairshare dominates while jobs are young, and it’s only if they’ve been pending for a long time that age really starts to overtake fair share.  We also set QoS priority weight very high, so that really critical jobs go straight to the top of the queue, but those qos’s are always tightly constrained to a very small number of resources (we have a ‘priority’ qos, but it only allows the user to consume 16 CPUs and a single GPU)

I have to say, I find this to be an area where SLURM is a bit weaker than some other schedulers.  It’s very difficult, sometimes, to really understand why a particular job isn’t running.  I used to be an LSF administrator, and I really loved the ‘bjobs -l -p’ command in LSF, which tells you exactly why a job cannot be run on each node, and the answer can be different in each case.

Tim

Yes, we've see the same thing with mosaic/heterogeneous partitions. Our solution is to split based on hardware type.

Having a bunch of partitions may seem unwieldy but the scheduler can handle it. For instance we have 110 partitions and the scheduler handles it fine (most of those are hardware owned by specific groups not public partitions everyone can see). We've taken up the convention of naming our partitions after the hardware type. For instance we have a gpu partition (our A100's) and a gpu_h200 partition. Making it easy for people to identify the hardware. People who can use both will leverage mutltipartition submission ala #SBATCH -p gpu,gpu_h200.

I don't know of a good solution if you want to keep the mosiac partition as it really requires you users to think at a higher level and realize there is vacant hardware that could be used if they just selected a different gpu type. Having a separate partition makes it much easier to see.

-Paul Edmon-

We have several different types of GPUs in the same 'gpu' partition.  The problem we're having occurs when one of those types of GPUs is fully occupied and there are a bunch of queued jobs waiting for those GPUs.  If someone requests idle GPUs of a different type, those jobs end up getting stalled, even though there are plenty of GPUs available.

For example, say we have 10 A100 GPUs and 10 H100 GPUs.  If there are 10 H100 GPU jobs running and more in queue waiting for them, subsequently submitted A100 jobs will sit in queue even if there are plenty of idle A100 GPUs.  The only way we can get the A100 jobs to run is by manually bumping their priority higher than the pending H100 jobs.

Has anyone else encountered this issue?  The only way we can think of to potentially solve it is to have separate partitions for each GPU type, but that seems unwieldy.

We are currently running Slurm 24.05.8.

Thanks,

Kevin

--

Kevin Hildebrand

Director of Research Technology and HPC Services

Division of IT