Slurm commands and environment variables

Slurm Scheduler  

Slurm  is the job scheduler deployed on JASMIN. It allows users to submit, monitor, and control jobs on the LOTUS (CPU) and ORCHID (GPU) clusters.

Essential Slurm commands  

Slurm command	Description
sbatch my_batch_script.sh	Submit a job script to the scheduler
sinfo	Show available scheduling queues
squeue --me	List my pending and running jobs
salloc -p debug -q debug -A mygws	Request an interactive session on LOTUS

Job specification  

Long and short argument names are separated by a comma.

#SBATCH  

• Scheduler directive - goes in front of the arguments below in a job script file
• An example Slurm job script file is available here

--account=GWS_NAME, -A GWS_NAME  

• Specify which project’s account to log the compute with by replacing GWS_NAME
• To choose the right one, please read about the new Slurm job accounting by project

--partition=PARTITION_NAME, -p PARTITION_NAME  

• Specify the scheduling partition by replacing PARTITION_NAME
• See the Slurm queues page for the list of partitions that you can use

--qos=QOS_NAME, -q QOS_NAME  

• Specify what Quality of Service your task needs by replacing QOS_NAME
• See the list of QoS that you can use

--time=hh:mm:ss, -t hh:mm:ss  

• Set the maximum runtime limit by replacing hh:mm:ss

--time-min=hh:mm:ss  

• Set an estimated runtime by replacing hh:mm:ss

--job-name=JOB_NAME  

• Specify a name for the job by replacing JOB_NAME

--output=FILE_NAME, -o FILE_NAME  

• Standard job output - where your program prints to normally (stdout)
• Defaults: appends to the file and file name is slurm-%j.out, where %j is replaced by the job ID

--error=FILE_NAME, -e FILE_NAME  

• Standard error output - where your program prints to if an error occurs (stderr)
• Defaults: appends to the file and file name is slurm-%j.err, where %j is replaced by the job ID

--open-mode=append|truncate  

• Write mode for error/output files
• Pick either append or truncate

--mem=XXX  

• Specify that XXX memory is required for the job. Default units are megabytes (e.g. --mem=250 means 250MB) but you can specify the unit, e.g. --mem=5G for 5 GB.

--array=INDEX  

• Specify a job array, e.g. --array=1-10 - for an example submission script, see this page
• The default standard output file name is slurm-%A_%a.out, where %A is replaced by the job ID and %a with the array index
• To change this, use --output and --error as above with %A and %a instead of %j

--array=INDEX%ArrayTaskThrottle  

• A maximum number of simultaneously running tasks from the job array may be specified using a % separator
• For example, --array=1-15%4 will limit the number of simultaneously running tasks from this job array to 4

--chdir=DIRECTORY, -D DIRECTORY  

• Set the working directory of the batch script to DIRECTORY before it is executed

--exclusive  

• Exclusive execution mode

--dependency=<dependency_list>  

• Defer the start of this job until the specified dependencies have been satisfied as completed
• See the Slurm documentation  for examples

--ntasks=NUMBER_OF_CORES, -n NUMBER_OF_CORES  

• Number of CPU cores

--constraint=HOST_GROUP_NAME  

• To select a node with a specific processor model
• A list of host groups that you can use are available here

Job control commands  

Slurm command	Description
scancel <jobid>	Kill a job
scontrol show job <jobid>	Show details job information
scontrol update job <jobid>	Modify a pending job
scancel --me	Kill all jobs owned by a user

Job environment variables  

Slurm variable	Description
$SLURM_JOBID	Job identifier number
$SLURM_ARRAY_JOB_ID	Job array
$SLURM_ARRAY_TASK_ID	Job array index
$SLURM_ARRAY_TASK_MAX	Last index number within a job array
$SLURM_NTASKS	Number of processors allocated