Useful Slurm Commands

Sample slurm script for many independent jobs

#!/bin/bash
#SBATCH --time=<desired time up to node upper time limit> #lower times get higher priority in the queuing system
#SBATCH --nodes=3
#SBATCH --ntasks=(# cores per node) * (# nodes) #this should equal the number of jobs you want to run at the same time
#SBATCH --job-name=<your_job_name>
#SBATCH --output=<std_out_filename>
#SBATCH -p <partition_name>  
 
DOCK_PATH="<path to dock bin/executable>"
JOB_FILE="<your common input file>"
CDIR=$(pwd)
# Assuming your paths to experiments are listed line by line in a file named paths.txt
while IFS= read -r path; do
  cp $JOB_FILE -t $path
  cd $path
  base=$(basename -s .in $JOB_FILE)
  # You can modify this srun command based on your requirements, the -n1 requests 1 core for the srun job, the -N1 requests 1 node, the --exclusive prevents this job's cores from being used in other jobs, -W 0 tells the script to not give   the job a timelimit
  srun --mem=6090 --exclusive -N1 -n1 -W 0 $DOCK_PATH/dock6.rdkit -i $base.in -o $base.out &
  cd $CDIR
done < paths.txt
wait #necessary to prevent the script from ending and terminating your jobs

Basic Slurm commands

• squeue -u $(whoami) - displays your running jobs and their job ids
• sinfo -l -a | grep idle - displays available nodes (note: some nodes may have exclusive access to certain lab groups)
• scancel <jobid>- remove the job from the queue

Use the man or info command in Unix to get more details of usage for these commands. You can use PBS commands inside your script. These are usually optional, but can be useful. There is an example script see NAMD on Seawulf

#PBS -l nodes=8:ppn=2
#PBS -l walltime=08:00:00
#PBS -N my_job_name
#PBS -M user@ic.sunysb.edu
#PBS -o namd.md01.out
#PBS -e namd.md01.err
#PBS -V

Using these PBS directives will name the job, the output files etc.

#PBS -j oe
#PBS -o pbs.out    

This will join the output and error streams into the output file. $PBS_O_WORKDIR is an environment variable that contains the path the script was submitted in. Usually you want to define a specific workdir and use that instead of relying on this variable.

Advanced Tricks

Delete all your jobs (either will work, check different versions of PBS)

qstat -u sudipto | awk -F. '/sudipto/{print $1}' | xargs qdel
qstat | grep sudipto | awk -F. '{print $1}' | xargs qdel 

Delete all your queued jobs only. Leaves all runnings jobs alone.

qstat -u sudipto | awk -F. '/00 Q /{printf $1" "}' | xargs qdel
qstat -u sudipto | awk -F. '/ Q   --/{printf $1" " }' | xargs qdel

List all working nodes in the queue

pbsnodes | egrep '^node|state =' | grep -B 1 'state = free' | grep ^node

set NODELIST=`pbsnodes | egrep '^node|state =' | grep -B 1 'state = free' | grep ^node`
foreach node ($NODELIST) 

/usr/local/torque-2.1.6/bin/qsub -l nodes=${NODE} ${HOME}/get.nodes.stats.csh done

Run on a particular type of node

 qsub -l nodes=1:beta:ppn=1 ${script}