GPUs, Constraints & Features

Beyond standard CPU/memory resources, the cluster provides special resources you can request via --gres and --constraint.

GPU Jobs

To request a GPU, use the gpu-general-pool partition and add --gres=gpu:1:

#!/bin/bash
#SBATCH --job-name=gpu_job
#SBATCH --account=my_account
#SBATCH --partition=gpu-general-pool
#SBATCH --qos=my_qos
#SBATCH --time=02:00:00
#SBATCH --ntasks=1
#SBATCH --nodes=1
#SBATCH --cpus-per-task=10
#SBATCH --gres=gpu:1
#SBATCH --mem-per-cpu=4G
#SBATCH --output=gpu_job_%j.out
#SBATCH --error=gpu_job_%j.err

module purge
module load python/python-3.8

# Your GPU commands here

Constraints

Constraints let you target nodes with specific hardware. Add --constraint=<feature> to your job script:

#SBATCH --constraint=localscratch,amd

Multiple constraints are comma-separated — the job will only run on nodes matching all of them.

Available Features

Run features on the login node to see the current list. Current features:

Using Local Scratch

If your job uses /localscratch, you must clean up after yourself — space is shared across all jobs on that node. Add this to your script:

export CACHEDIR=/localscratch/${USER}_${SLURM_JOB_ID}
mkdir -p $CACHEDIR

cleanup() {
  rm -rf -- "$CACHEDIR" || true
}
trap cleanup EXIT INT TERM HUP

The trap ensures cleanup runs even if the job fails or is cancelled.

Available GRES Types

GresTypes=gpu,amd,af3,intel,localscratch