Job submission

Slurm job’s terminology: job, job step, task and CPUs

A slurm job (submitted via sbatch) can consists of multiple steps in series. Each step (specified via srun) can run multiple tasks (ie programs) in parallel. Each task gets its own set of CPUs. As an example, consider the workflow and corresponding breakdown shown in fig 2.

In this example, note:

• When you explicitly request 1 CPU per task (--cpus-per-task=1), you should also explicitly specify the number of tasks (--ntasks). Otherwise, srun may start the task twice in parallel (because CPUs are allocated in multiples of 2)
• The default slurm allocation is a single task and single CPU (ie --ntasks=1 --cpus-per-task=1). Thus, it is not necessary to explicitly request these to run a single task on a single CPU.
• When using multiple tasks, specify --mem-per-cpu.