SLURM examples

SLURM.md

Note: this page was originally written in pmWiki for the wiki of the Biomedia group.

It was converted to markdown using an HTML-to-markdown converter and a markdown editor.

See the official documentation for more info about SLURM.

Example 1: looping over subjects

the --wrap option allows you to pass one-liners to sbatch (see sbatch -h):

--wrap[=command string] wrap commmand string in a sh script and submit

It can be used to call sbatch for each subject within a for loop:

for i in {0..54}
do 
  sbatch --mem=5G -c 10 --wrap="python leaveoneout_segmentation.py $i --n_jobs=10"
done

Alternatively, this could have been done with arrays: sbatch --array=0-55 and then relying on the SLURM_ARRAY_TASK_ID environment to retrieve your parameter value.

Example 2: calling sbatch from Python

You can use a function such as:

import subprocess

__all__ = [ "sbatch" ]

def sbatch( cmd, mem=5, c=10, verbose=False, dryrun=False ):
    sbatch_cmd = 'sbatch --mem=' + str(mem) + 'G -c ' + str(c) + ' --wrap="' + cmd + '"'

    if verbose or dryrun:
        print sbatch_cmd

    if dryrun:
        return

Example 3: handy bash functions

Launch an interactive session through SLURM:

function ssh_slurm {
    ntasks=${1:-8}
    mem=${2:-32}
    ssh [email protected] -t "salloc -p interactive -c${ntasks} --mem=${mem}G srun -c${ntasks} --mem=${mem}G --pty /bin/bash"
 }

Launch the ipython notebook web server, retrieve the name of the machine assigned to it, setup an ssh tunnel and finally open a web browser:

function notebook {
    folder=${1:-/vol/medic02/users/kpk09}
    ncpus=${2:-8}
    ntasks=${3:-10}
    mem=${4:-32}
    port1=${5:-8888}
    port2=${6:-8890}

    # beware of [ -z "$PS1" ] && return at the beginning of your .bashrc
    # (hence .bash_setup)
    # see [https://github.com/ipython/ipython/issues/2426/](https://github.com/ipython/ipython/issues/2426/)
    # for IPYTHONDIR and the crash on a network folder
    job_id=`echo "source ~/.bash_setup && sbatch -p interactive --mem=${mem}G -c${ncpus} --wrap='cd $folder && export IPYTHONDIR=/tmp && ipython notebook --no-browser --port=${port1}'" | ssh  [email protected] "/bin/bash -s" | perl -ne '@a =split(" ");print \$a[-1]'`

    if [ -z "${job_id}" ]; then
        echo "sbatch failed"
        return
    fi

    echo "job id: ${job_id}"

    # sleep to give ipython the time to get started
    sleep 10

    node_id=`ssh [email protected] "squeue -u kpk09 -o %N -j ${job_id}" | tail -1`

    if [ -z "${node_id}" ]; then
        echo "could not obtain node id"
        return
    fi

    echo "node id: ${node_id}:${port1}"

    # setup SSH tunnel
    ssh -N -f -L localhost:${port2}:localhost:${port1} kpk09@${node_id}.doc.ic.ac.uk

    echo "url: [http://localhost:$](http://localhost:$){port2}/"

    chromium-browser [http://localhost:$](http://localhost:$){port2}/
}

On a Mac, you can replace the last line by:

open -a Safari [http://localhost:$](http://localhost:$){port2}/

Run scancel with or without argument:

function scancel {
    ssh [email protected] "scancel $@" 
}

Simple wrapper for the whole command line:

function mysbatch {
    sbatch --mem=5G -c 10 --wrap="$*"
}

Example 4: reassign pending jobs to a different partition

for i in `squeue -u kpk09 | grep PD | perl -ne 's/\s+/ /g; @a=split(" ",$_); print $a[0]." ";'`;
do 
  scontrol update JobId=$i Partition=long;
done