GitLab

Here the suggestions from a discussion with Dmitry Ozerov:

1- prepare a single slurm script containing an array of jobs

ex: sbatch -a 1-100 test.sh

this will submit in one command 100 jobs to slurm. The trick to be nice with the other users:

sbatch -a 1-100%10 test.sh

(or similar %25, %50)

2- nodes and tasks:

for ntasks < 16 than it would be more meaningful to ensure that the ntasks run on the same node. Now:

#SBATCH --ntasks=4

#SBATCH --ntasks-per-node=4

suggested:

#SBATCH --ntasks=4

#SBATCH --nodes=1

3- Loading modules in slurm script

module load gcc.....

so everything what is needed to load, setup to run the job (otherwise it's a dependence on what there is in .bashrc or environment (from which the job is submitted).