SLURM #SBATCH directives (quick reference)

SLURM #SBATCH directives (quick reference)

This reference is intentionally short and practical. SLURM policies vary by site; always check your cluster documentation.

Core directives

• #SBATCH --job-name=<name>: Human-friendly name shown in squeue.
• #SBATCH --time=<time>: Walltime limit (common portable formats: HH:MM:SS or D-HH:MM:SS; SLURM also accepts minutes-only and other variants).
• #SBATCH --partition=<partition>: Queue/partition (if required).
• #SBATCH --account=<account>: Project/account (if required).
• #SBATCH --nodes=<N>: Number of nodes.
• #SBATCH --ntasks=<R>: Total MPI ranks/tasks.
• #SBATCH --ntasks-per-node=<rpn>: Ranks per node (if --ntasks is not set, total ranks is typically nodes * rpn).
• #SBATCH --cpus-per-task=<t>: CPU cores per rank (commonly used for OpenMP threads).
• #SBATCH --mem=<size>: Memory per node (integer MB by default, or with suffix K/M/G/T, e.g. 32G; 0 often means “all memory on node”).
• #SBATCH --mem-per-cpu=<size>: Memory per CPU core (same units as --mem).

GPUs

GPU directives differ across clusters. A common portable pattern is:

• #SBATCH --gres=gpu:<count> (or gpu:<type>:<count>)

Examples:

• #SBATCH --gres=gpu:4
• #SBATCH --gres=gpu:a100:4

Output and diagnostics

• #SBATCH --output=slurm-%j.out: Stdout file (%j is job id).
• #SBATCH --error=slurm-%j.err: Stderr file.

Inside the script, useful diagnostics:

echo "job_id=$SLURM_JOB_ID"
echo "nodes=$SLURM_JOB_NODELIST"
echo "submit_dir=$SLURM_SUBMIT_DIR"

Common mapping advice

• MPI-only: cpus-per-task=1, increase ntasks.
• Hybrid MPI+OpenMP: set cpus-per-task = OMP_NUM_THREADS, reduce ntasks.
• Prefer srun inside SLURM allocations; it integrates with SLURM’s task placement.