Anatomy of a SLURM Batch Job

The following is an example of a regular SLURM Batch Job:

The top half of this script sets up SLURM. Crucially, it specifies the maximum amount of time to allocate to the run and the computing resources to allocate. You'll need to know how to set these to work with SLURM.

Time: This is specified as wall time: which is time as it passes on the clock. The alternative is CPU time, which is "wall time * the number of requested CPUs (SLURM calls this ntasks)". You really only need to specify one kind of time. The following command, #SBATCH --time=0:05:00, specifies 5 minutes of job time:

Computing Resources: Here you specify the number of nodes and CPUs (ntasks). Each of the main clusters (e.g., Puma, Ocelote etc.) has a certain number of nodes and each node has a limited number of CPUs and memory. So, if you ask for one node on Puma, you can have up to 94 CPUs and 5 GB of RAM per CPU. However, on Ocelote, you can have up to 28 CPUs for one node, but you get 6 GB of RAM for each CPU. Check the resources below for a table of Cluster and their associated nodes. The following command #SBATCH --nodes=1 --ntasks=1, specifies that you are requesting one node (you'll always need at least one) and one CPU (again, you'll always need at least 1).

HPC 5: Batch jobs