Lowpriority

From NJIT-ARCS HPC Wiki
Jump to: navigation, search

The "lowpriority" partition allows you to borrow reserved (private) nodes, including GPUs. There are caveats:

  • The jobs are immediately terminated (preempted) if the owner or owner's group submits a job that requires the resources used by the lowpriority job(s).
  • You must specify a runtime; the maximum runtime is three days.
  • You must checkpoint your job, please contact HPC@NJIT.EDU for assistance.
  • If your job is preempted or runs out of time, it will automatically be requeued.
  • Interactive jobs should not be run lowpriority as the jobs could be preempted without warning and no chance to save your work.

To utilize lowpriority, add these specifications:

#SBATCH --partition=lowpriority
#SBATCH --time 0-0:30:0 # Max 3 days
#SBATCH --constraint=avx512

You can specify GPU(s):

#SBATCH --gres=gpu:1

Alternatively, you can specify particular model (Type) of GPU(s):

#SBATCH --gres=gpu:Type:1 # See /etc/slurm/gres.conf for Type values