Jobs Categories and Limits
MCMeSU supercomputer resources are managed by Slurm scheduler.
Users must submit their jobs on one of the computing system according to their needs.
This page describes the jobs categories and the limits enforced on each system in order maximize resource occupation and minimize wait times. These limits can change dynamically according to MCMeSU load and the number of users or jobs waiting for computing resources. When a user has jobs pending with “QOSJobLimit” reason, that means it has reached the limit of the corresponding job category.
Users do not have to declare their jobs category explicitly, it is derived from their resources needs (ncpus, walltime, etc.).
Overall, remember that resources are managed in a fair-share fashion, meaning that users that compute the most have a lower priority as long as other users with lower historic usage of the machine have jobs pending.
In order to specify the walltime limit of a job, use the --time=D-HH:MM:SS parameter.
For example :
#SBATCH --time=4-03:20:10
Sets a maximal execution time of 4 days, 3 hours, 20 minutes and 10 seconds.
Jobs Categories and Limits
When submitting a job, use preferably a number of CPUs that is a multiple of 32 which corresponds to the number of CPUs available on each node.
| PARTITION | CATEGORY | WALLTIME LIMIT | OCCUPATION LIMIT |
|---|---|---|---|
| thin | Normal | 1 day | |
| thin | Long | 8 days | 4 nodes per user |
| std | Normal | 1 day | |
| std | Long | 8 days | 4 nodes per user |
| bigm | Normal | 1 day | |
| bigm | Long | 8 days | 1 node per user |
| gpu | Normal | 1 day | |
| gpu | Long | 8 days | 2 GPUs per user |
Note on GPU Resources
When submitting a job in the gpu partition, users should specify the number of GPUs they whish to use.
This is done by adding the --gpu= to the sbatch command line.
For example :
sbatch --gpu=2 -n 32 myjob.sh
will allocate your job 2 GPUs and 32 CPUs.
For the billing of CPU hours, the occupation of a GPU corresponds to 16 CPUs.