Slurm SBATCH Directive Reference Sheet

Directive	Purpose	Typical values / examples	Notes & tips
#SBATCH -J / --job-name=	Name the job	-J myjob	Shows in squeue, logs, emails. Keep it short.
#SBATCH -p / --partition=	Choose queue/partition	-p LocalQ	Controls where the job can run; ask your cluster which partitions exist.
#SBATCH -t / --time=	Wall time limit	-t 00:30:00 (hh:mm:ss)	Jobs are killed at limit; request slightly more than you expect.
#SBATCH -N / --nodes=	Number of nodes	-N 1	Multi-node jobs often need MPI; otherwise keep to 1.
#SBATCH -n / --ntasks=	Total parallel tasks	-n 16	For MPI or multiple srun tasks. Often paired with --cpus-per-task.
#SBATCH -c / --cpus-per-task=	Threads per task	-c 8	For OpenMP/numba/BLAS-threaded codes. Set OMP_NUM_THREADS=$SLURM_CPUS_PER_TASK.
#SBATCH --ntasks-per-node=	Tasks per node	--ntasks-per-node=4	Helpful for multi-node placement control.
#SBATCH --mem=	Memory per node	--mem=16G	Mutually exclusive with --mem-per-cpu. Covers the whole node allocation.
#SBATCH --mem-per-cpu=	Memory per CPU	--mem-per-cpu=2G	Multiplies by total CPUs allocated. Good when scaling tasks.
#SBATCH -o / --output=	STDOUT file	-o OUT/%x_%j.out	%x=job name, %j=job ID. Use directories you can write to.
#SBATCH -e / --error=	STDERR file	-e OUT/%x_%j.err	Keep separate from --output for easier debugging.
#SBATCH -D / --chdir=	Set working dir	-D /path/to/run	Equivalent to cd before job starts.
#SBATCH --mail-type=	Email on events	--mail-type=BEGIN,END,FAIL	Other options: ALL, TIME_LIMIT, REQUEUE, etc.
#SBATCH --mail-user=	Email recipient	--mail-user=you@uni.edu	Must be a valid address; some clusters default to your account email.
#SBATCH -A / --account=	Charge account	-A proj123	Required on centers with allocations.
#SBATCH --qos=	Quality of service	--qos=normal	Affects limits/priority if your site uses QoS.
#SBATCH --array=	Job arrays	--array=0-99%10	%10 throttles concurrency. Inside script use $SLURM_ARRAY_TASK_ID.
#SBATCH -d / --dependency=	Job dependencies	-d afterok:12345	Variants: after, afterok, afternotok, afterany, singleton.
#SBATCH --gres=	Generic resources (GPUs)	--gres=gpu:1	Cluster-specific (e.g., gpu:tesla:2). Often requires GPU partition.
#SBATCH --constraint=	Node features	--constraint=rome&256G	Matches nodes with given features (labels vary by site).
#SBATCH --exclusive	Exclusive node use	--exclusive	You get the whole node(s); don’t combine with small --mem requests.
#SBATCH --hint=	SMT/NUMA hint	--hint=nomultithread	May improve performance by disabling hyper-threads.
#SBATCH --export=	Env var export	--export=ALL,MYFLAG=1	NONE to block, or enumerate required vars explicitly.
#SBATCH --signal=	Send signal before kill	--signal=B:USR1@60	Send SIGUSR1 60s before preemption/time limit; add checkpoints.
#SBATCH --requeue / --no-requeue	Requeue on preempt/fail	--requeue	Needs --signal + checkpointing to be useful.
#SBATCH --open-mode=	Output file mode	--open-mode=append	truncate (default) overwrites; append extends existing logs.
#SBATCH --wrap=	One-liner job	--wrap="python script.py"	Handy for quick tests without a script file.
#SBATCH --reservation=	Use reservation	--reservation=workshop	For scheduled training events or maintenance windows.
#SBATCH --profile=	Perf profiling	--profile=task	Site-dependent; may emit stats to slurm.profiling files.
#SBATCH --nice=	Lower priority	--nice=1000	Useful when queue-friendly is preferred over speed.
#SBATCH --begin=	Deferred start	--begin=18:30 or now+2hours	Reserves spot to start at/after given time.
#SBATCH --hold / --nohold	Submit held	--hold	Release later with scontrol release <jobid>.
#SBATCH --time-min=	Flexible runtime	--time-min=00:10:00	Lets backfill with shorter slot if full time not available.
#SBATCH --tmp=	Local scratch	--tmp=40G	Requests node-local tmp space (if configured).