Advanced Job Submission
This is not a complete article: This is a Draft, a work in progress that is intended to be published into an article, which may or may not be ready for inclusion in the main wiki. It should not necessarily be considered factual or authoritative.
Submitting Numerous Compute Tasks
In case you have multiple data files to process with many different combinations of parameters, you should not have to create and submit thousands of job scripts to the scheduler. Here are three proposed tools to answer this kind of need:
- GNU Parallel: fill an entire compute node with serial compute tasks addressing multiple combinations of arguments.
- Job Arrays: a single script that represents multiple Slurm jobs, each of them identified by an integer value.
- GLOST: the Greedy Launcher Of Small Tasks is a kind of sophisticated mix of both above tools.
While Slurm jobs are building-blocks of pipelines, inter-job dependencies are the links and relationships between each step of pipelines. For example, if two different jobs need to run one after the other, the second job depends on the first one. The dependency could depend on the start time, the end time or the final status of the first job. Typically, we want the second job to be started only once the first job has succeeded:
JOBID1=$(sbatch --parsable job1.sh) # Save the first job ID sbatch --dependency=afterok:$JOBID1 job2.sh # Submit a job with a dependency to the first job
- Multiple jobs can have the same dependency (multiple jobs are waiting after one job)
- A job can have multiple dependencies (one job is waiting after multiple jobs)
- There are multiple types of dependencies:
afternotok, etc. For more details, please look for the
--dependencyoption on the official Sbatch documentation page.
The Slurm scheduler supports heterogeneous jobs. This could be very useful if you know in advance that your MPI application will require more cores and more memory for the main process than for other processes.
For example, if the main process requires 5 cores and 16GB of RAM, while other processes only require 1 core and 1GB of RAM, we can specify both types of requirements in a job script:
#!/bin/bash #SBATCH --ntasks=1 --cpus-per-task=5 --mem-per-cpu=16000M #SBATCH hetjob #SBATCH --ntasks=15 --cpus-per-task=1 --mem-per-cpu=1000M srun application.exe
Or we can separate resource requests with a colon (
:) on the
sbatch commande line:
sbatch --ntasks=1 --cpus-per-task=5 --mem-per-cpu=16000M : --ntasks=15 --cpus-per-task=1 --mem-per-cpu=1000M mpi_job.sh