Open ptheywood opened 2 years ago
Bede's usage documentation on filesystems currently only mentions the NFS/lustre mounts /projects, /nobackup and /users.
/projects
/nobackup
/users
Making users aware of scratch / per node temporary storage can be benficial to job and NFS/lustre performance.
E.g. the Sheffield HPC documentation of this
It will be worth enquiring about this with the sysadmins to ensure this is meaningful / correct for bede.
On login nodes, $TMPDIR is not set, so /tmp is likely used. Within slurm jobs, $TMPDIR is set to /tmp.
$TMPDIR
/tmp
/tmp on gpu nodes is a 1.6T raid1 partition split across the 2 SSDs in the node.
gpu
$SLURM_JOB_ID contains the integer mapping of the slurm job id, which can be used for per-job temp files combined with TMPDIR.
$SLURM_JOB_ID
If the slurm configuration were changed, then using this would not cause issues (just extra nesting).
I.e. something vaguely like
# slurm options... #$ ... # Create the temp directory. mkdir -p ${TMPDIR}/${SLURM_JOB_ID} # Reset $TMPDIR to be more specifc OLD_TMPDIR=${TMPDIR} TMPDIR=${TMPDIR}/${SLURM_JOB_ID} # Do stuff # Copy out results mv ${TMPDIR}/path/to/source ${HOME}/path/to/dest # Cleanup rm -rf ${TMPDIR} # Unset TMPDIR TMPDIR=${OLD_TMPDIR}
Bede's usage documentation on filesystems currently only mentions the NFS/lustre mounts
/projects
,/nobackup
and/users
.Making users aware of scratch / per node temporary storage can be benficial to job and NFS/lustre performance.
E.g. the Sheffield HPC documentation of this
It will be worth enquiring about this with the sysadmins to ensure this is meaningful / correct for bede.
On login nodes,
$TMPDIR
is not set, so/tmp
is likely used. Within slurm jobs,$TMPDIR
is set to/tmp
./tmp
ongpu
nodes is a 1.6T raid1 partition split across the 2 SSDs in the node.$SLURM_JOB_ID
contains the integer mapping of the slurm job id, which can be used for per-job temp files combined with TMPDIR.If the slurm configuration were changed, then using this would not cause issues (just extra nesting).
I.e. something vaguely like