Partition Name
Use Cases
priority
  • Treat as default CPU queue
  • Long running, non-interruptible jobs
  • When cluster is heavily utilized (minimal wait time)
unsafe
  • Run more jobs than priority but could be interrupted
  • Good when the cluster in not heavily utilized
  • Smaller / shorter jobs
  • Higher throughput
  • Job arrays (lots of small jobs)
  • Resumable jobs
gpupriority
  • Treat as default GPU queue
  • Long running, non-interruptible jobs
  • When cluster is heavily utilized (minimal wait time)
  • A40 < A100 for compute
  • A40 45GB > A100 40GB VRAM
gpuunsafe
  • Run more jobs than gpupriority but could be interrupted
  • Good when most GPUs are not allocated
  • Smaller / shorter jobs
  • Higher throughput
fairshare
  • Large jobs that use many CPUs
    • Big MPI jobs
  • Large jobs with high memory usage
gpufairshare
  • Multi-GPU jobs are much faster
    • Up to 4 GPUs at once
  • AI/ML workloads
  • Jobs that use many CPUs, lots of RAM, and GPUs
  • GPU jobs that can be checkpointed