Running Slurm Jobs on Hex
Rather than reiterate the contents of the absolutely excellent online documentation for Slurm over at the SchedMD pages, we instead link to these guides here:
If you have a particularly good source for more Slurm or related batch compute technology, please do contact us so we can update this list!
Partitions on Hex #
Rather than forcing users to fully specify their resource needs on each run, Hex instead provides a number of partition queues that describe particular environments for you.
These can be found by running sinfo on the login node, and follow the general schema: [type]-[duration] whereby type encodes the runtime's main hardware resource and duration encodes the maximum runtime for a job on this resource.
Currently (June 2026) we have a2000, a5000 and cpu job types, which map to our RTX A2000 GPUs, RTX A5000 GPUs, and regular CPU-only workers respectively.
Our durations are 48h, 6h and 5m options for each of the above; mapping to 48-hours maximum, 6-hour maximum and 5-minute maximum runtimes.
These durations should be used as a 'safety net' for processes running on Hex, where if you are looking to do short test runs, placing the jobs on the *-5m queues will place the jobs at a slightly higher priority than the longer runs to allow a consistent quality-of-service across the cluster for development. the *-6h and *-48h queues have correspondingly lower priority so equipment is not needlessly tied up on excessive runtimes.
Note that these are limits for runtime, your processes can quit early and the resources will be immediately released to the cluster for other jobs.