Job submission options
#SBATCH directives are Slurm options embedded in a batch script as special comments. They are equivalent to passing the same flags on the sbatch command line. The full reference is in the Slurm sbatch documentation, but the most commonly used options are described below.
Most of these options also apply to salloc and srun when passed directly on the command line, though some options are specific to batch submission.
Directives must come before any executable line
Slurm stops parsing #SBATCH directives as soon as it encounters the first executable line (any line that is not a comment or blank). Directives placed after executable code are silently ignored. Always place all #SBATCH lines at the top of the script, right after the shebang.
Example script#
A typical batch job using the most common options:
job.sbatch
#!/bin/bash
#SBATCH --job-name=analysis # job name (shown in squeue and notifications)
#SBATCH --output=%x.%j.out # stdout file (%x = job name, %j = job ID)
#SBATCH --error=%x.%j.err # stderr file
#SBATCH --time=4:00:00 # wall-clock time limit (HH:MM:SS)
#SBATCH --partition=normal # partition to submit to
#SBATCH --cpus-per-task=4 # CPU cores per task
#SBATCH --mem=16GB # memory per node
#SBATCH --mail-type=END,FAIL # send email on job end or failure
module load python/3.11.7
python3 analysis.py
Common #SBATCH options#
Resources#
| Option | Short | Description | Example |
|---|---|---|---|
--cpus-per-task=<n> | -c | Number of CPU cores per task. Set to 1 unless your code is explicitly multi-threaded or uses OpenMP. | #SBATCH --cpus-per-task=8 |
--mem=<size> | Total memory per node. Accepts K, M, G, T suffixes. | #SBATCH --mem=16GB | |
--mem-per-cpu=<size> | Memory per allocated CPU core. Alternative to --mem; do not use both. | #SBATCH --mem-per-cpu=4GB | |
--ntasks=<n> | -n | Number of parallel tasks (MPI ranks). Defaults to 1. | #SBATCH --ntasks=4 |
--nodes=<n> | -N | Number of nodes to allocate. | #SBATCH --nodes=2 |
--gpus=<n> | -G | Number of GPUs to allocate. Requires submitting to the gpu partition. | #SBATCH --gpus=1 |
--time=<time> | -t | Wall-clock time limit. Accepted formats: MM:SS, HH:MM:SS, D-HH:MM:SS. The job is killed when the limit is reached. | #SBATCH --time=2:00:00 |
--partition=<name> | -p | Partition to submit to (e.g. normal, gpu, bigmem, owners). | #SBATCH --partition=gpu |
Output#
| Option | Short | Description | Example |
|---|---|---|---|
--job-name=<name> | -J | Name shown in squeue, sacct, and email notifications. | #SBATCH --job-name=my_run |
--output=<file> | -o | File for standard output. %j expands to the job ID, %x to the job name. Defaults to slurm-%j.out. | #SBATCH --output=%x.%j.out |
--error=<file> | -e | File for standard error. If omitted, stderr is merged into the --output file. | #SBATCH --error=%x.%j.err |
--open-mode=append | Append to output files instead of overwriting them. Useful for recurring jobs. | #SBATCH --open-mode=append |
Notifications#
| Option | Short | Description | Example |
|---|---|---|---|
--mail-user=<addr> | Email address for notifications. Defaults to your SUNet address. | #SBATCH --mail-user=you@stanford.edu | |
--mail-type=<events> | Comma-separated list of events that trigger an email. Common values: BEGIN, END, FAIL, ALL. | #SBATCH --mail-type=END,FAIL |
Job control#
| Option | Short | Description | Example |
|---|---|---|---|
--dependency=<type:jobid> | -d | Delay the job until a condition on another job is met. Common types: afterok (start after successful completion), afterany (start regardless of exit status), singleton (only one job with this name runs at a time). See the Slurm dependency documentation for the full list. | #SBATCH --dependency=afterok:12345 |
--array=<range> | -a | Submit a job array. Each element runs the same script with a unique $SLURM_ARRAY_TASK_ID. Ranges can include step sizes and a maximum number of simultaneously running tasks. | #SBATCH --array=1-100%10 |
--begin=<time> | -b | Defer job start until a given time. Accepts absolute timestamps or relative offsets like now+1hour, now+7days. | #SBATCH --begin=now+2hour |
--signal=<sig>@<n> | Send signal <sig> to the job <n> seconds before reaching the time limit. Useful to trigger graceful shutdown or resubmission in persistent jobs. | #SBATCH --signal=B:SIGUSR1@90 |
Node selection#
| Option | Short | Description | Example |
|---|---|---|---|
--constraint=<feature> | -C | Require nodes with a specific feature (e.g. GPU model, CPU generation). See the Node features page for available values. | #SBATCH --constraint=GPU_SKU:A100_SXM4 |