Controlling slurm jobs
Settings for slurm job
| Command | Effect |
|---|---|
| #SBATCH --account=staff | user is part of this group |
| #SBATCH --job-name=lvl_job | name of MY job |
| #SBATCH --gpus-per-node=1 | 1 GPU on worker node (default) |
| #SBATCH --output=user.log | output log file, if needed |
| #SBATCH --mem-per-cpu=3G | 3 gb memory per process |
| #SBATCH --ntasks=4 | you want 4 processes |
| #SBATCH --cpus-per-task=2 | 2 cores per process |
| #SBATCH --nodes=1-3 | use at LEAST one node,up to 3 |
| #SBATCH -- | w |
The default memory a batch jobs gets is 4gb, if user doesn't specify any setting in the batch file [DefMemPerNode=4096]
Partitions/queue config
The slurm cluster has 3 partitions (queues)
doTrain is (default), only for staff, allows jobs to run infinitely long time
All partition have DefaultTime=04:00:00
on allWork the MaxTime is 7 days and 1 hour
on beQuick the MaxTime is 1 days and 12 hour
Node config
NodeName=A … TmpDisk=1536
TmpDisk : Total size of temporary disk storage in TmpFS in megabytes
MinTmpDiskNode
default value for TmpDisk is 0, local scratch amount of TmpDisk space must be defined in the node config