Slurm Usage: A Primer

Slurm Scripts

Slurm scripts are text files you will need to create in order to submit a job to the scheduler. Slurm scripts start with #!/bin/bash (with optional flags) and contain a set of directives (which start with #SBATCH), followed by commands (at least one of which should start with srun):

#!/bin/bash -e
#SBATCH --job-name=JobName # job name (shows up in the queue) #SBATCH --account=nesi99999 # Project Account #SBATCH --time=00:10:00 # Walltime (HH:MM:SS) #SBATCH --mem-per-cpu=4096 # memory/cpu (in MB) #SBATCH --ntasks=2 # number of tasks (e.g. MPI) #SBATCH --cpus-per-task=4 # number of cores per task (e.g. OpenMP) #SBATCH --partition=long # specify a partition #SBATCH --hint=nomultithread # don't use hyperthreading
#SBATCH --output=%x-%j.out # %x and %j are replaced by job name and ID
#SBATCH --error=%x-%j.err
#SBATCH --mail-type=ALL # Optional: Send email notifications
#SBATCH # Use with --mail-type option srun [options] <executable> [options]

We strongly recommend using #!/bin/bash -e instead of plain #!/bin/bash, so that a command throwing an error will cause your job to stop, instead of wasting your project's CPU core hours by continuing to make use of potentially erroneous intermediate data.

Not all directives need to be specified, just the ones you need.

Launching executables with srun

The srun command runs the executable along with its options, within the resources allocated to the job. Although srun can be used from the command line in certain circumstances, it is best used within a Slurm job script, to be executed when the job is run by the scheduler. 

For MPI jobs, srun sets up the MPI runtime environment needed to run the parallel program, launching it on multiple CPUs, which can be on different nodes. srun should be used in place of any other MPI launcher, such as aprun or mpirun.

Commonly Used Slurm Environment variables

These can be useful within Slurm scripts:

  • $SLURM_JOB_ID (job id)
  • $SLURM_NNODES (number of nodes)
  • $SLURM_NTASKS (number of MPI tasks)
  • $SLURM_CPUS_PER_TASK (CPUs per MPI task)
  • $SLURM_SUBMIT_DIR (directory job was submitted from)
  • $SLURM_ARRAY_JOB_ID (job id for the array)
  • $SLURM_ARRAY_TASK_ID (job array index value)

MPI Jobs

For MPI jobs you need to set --ntasks to a value larger than 1, or if you want all nodes to run the same number of tasks, set --ntasks-per-node and --nodes instead.

OpenMP Jobs

For OpenMP jobs you need to set --cpus-per-task to a value larger than 1. Our Slurm prolog will then set OMP_NUM_THREADS to equal that number. Along with --cpus-per-task, you should set --ntasks (or --ntasks-per-node) to ensure that threading behaves correctly. For a simple OpenMP job (that doesn't also use MPI), --ntasks=1 should suffice.

Submitting a job

Use sbatch <script> to submit the job. All Slurm directives can alternatively be specified at the command line, e.g. sbatch --account=nesi12345 <script>. This overwrites directives specified in the script.

Try submitting a simple job

Submit job

#!/bin/bash -e
#SBATCH --job-name=hello
#SBATCH --time=00:02:00

srun echo "Hello, World!"

with sbatch --account=nesi12345 where nesi12345 is your NeSI project’s code. If you only have one project then you don’t need to specify it.

Submitting a job using GPGPU nodes

To submit to the general purpose GPU nodes, you need to add the following to your SLURM script:

#SBATCH -p gpu
#SBATCH --gres=gpu

Submitting a job between Māui and Māui_Ancil

Māui consists of the XC50 and the CS500 (Māui_Ancil) part. To submit a job from the XC50 part (including Māui login nodes) to the CS500 part you need to add:

#SBATCH --clusters=maui_ancil 
#SBATCH --export=NONE

Thus a prepost job submitted to the CS500 nodes from the Maui login node would look like:

#!/bin/bash -e
#SBATCH --job-name=hello
#SBATCH --time=00:02:00
#SBATCH --clusters=maui_ancil
#SBATCH --export=NONE

module load Anaconda2

Note: the --clusters need to be also specified for the other slurm tools to monitor other parts.

Monitor jobs

You can use squeue -u $USER to monitor your job status. Alternatively you can also use sview.

Checking completed jobs with sacct

Another useful Slurm command is sacct which retrieves information about completed jobs. For example:

sacct -j 14309

where the argument passed to -j is the job ID, will show us something like:

       JobID    JobName  Partition    Account  AllocCPUS      State ExitCode
------------ ---------- ---------- ---------- ---------- ---------- --------
14309       NeSI  nesi99999         80  COMPLETED      0:0
14309.batch       batch             nesi99999         80  COMPLETED      0:0
14309.0         yourapp             nesi99999         80  COMPLETED      0:0

By default sacct will list all of your jobs which were (or are) running on the current day. Each job will show as more than one line (unless -X is specified): an initial line for the job as a whole, and then an additional line for each job step, i.e.: the batch process which is your executing script, and then each of the srun commands it executes.

By changing the displayed columns you can gain information about the CPU and memory utilisation of the job, for example

sacct -j 14309 --format=jobid,jobname,elapsed,avecpu,totalcpu,alloccpus,maxrss,state
      JobID    JobName    Elapsed     AveCPU   TotalCPU  AllocCPUS     MaxRSS      State
------------ ---------- ---------- ---------- ---------- ---------- ---------- ----------
14309   00:12:42             00:00.012         80             COMPLETED
14309.batch       batch   00:12:42   00:00:00  00:00.012         80      1488K  COMPLETED
14309.0         yourapp   00:12:41   00:12:03   16:00:03         80    478356K  COMPLETE


Was this article helpful?
0 out of 0 found this helpful