Revision as of 09:53, 3 July 2018

Introduction

Overview

Viper uses the Slurm job scheduler to provide access to compute resource. The scheduler

A submission script is a file that provides information to Slurm about the task you are running so that it can be allocated to the appropriate resource, then sets up the environment so the task can run. A minimal submission script has three main components:

• A set of directives, staring with #SBATCH, which tell the scheduler about the job such as information about resource required, job name and job log and error files. Although in a normal BASH script anything starting with a '#' would indicate a comment, however the SLURM interpreter will recognise this as a command to pass to the scheduler.
• Information about how the job environment should be set up, for example what application modules should be loaded.
• The actual command(s) that need to be run.

SLURM script

The script starts with #!/bin/bash (also called a shebang), which makes the submission script a Linux bash script.

The script continues with a series of lines starting with #, which represent bash script comments. For Slurm, the lines starting with #SBATCH are directives that request job scheduling resources. (Note: it's very important that you put all the directives at the top of a script, before any other commands; any #SBATCH directive coming after a bash script command is ignored!)

The resource request #SBATCH --nodes=n determines how many compute nodes a job are allocated by the scheduler; only 1 node is allocated for this job. A note of caution is on threaded single process applications (e.g. Matlab). These applications cannot run on more than a single compute node; allocating more (e.g. #SBATCH --nodes=2) will end up with the first node being busy and the rest idle.

The maximum walltime is specified by #SBATCH --time=T, where T has format h:m:s. Normally, a job is expected to finish before the specified maximum walltime. After the walltime reaches the maximum, the job terminates regardless whether the job processes are still running or not.

The name of the job can be specified too with #SBATCH --job-name="name".

Note : Linux and Slurm do not care about the name used for a submission script, however for ease of support, we would recommend you call your submission script a relevant name with a .job suffix, for example MATLABtest.job

Example batch scripts

The following are a set of basic job submission scripts along with relevant additional information on how these can be adjusted to suit the task.

To submit a job, use the sbatch jobscript.job for example:

[username@login01 ~]$ sbatch MATLABtest.job
Submitted batch job 289522

Basic batch job

#!/bin/bash
#SBATCH -J jobname          # Job name, you can change it to whatever you want
#SBATCH -n 1                # Number of cores 
#SBATCH -o %N.%j.out        # Standard output will be written here
#SBATCH -e %N.%j.err        # Standard error will be written here
#SBATCH -p compute          # Slurm partition, where you want the job to be queued 
#SBATCH --time=2:00:00      #   Set max wallclock time this case 2 days
 
module purge
module add modulename
 
command

• Setting the time with #SBATCH --time=DD:HH:MM (or -t DD:HH:MM) allows the scheduler to place your job more efficiently, using the default here may cause delays when the cluster is running at a high demand rate.

Exclusive batch job

#!/bin/bash
#SBATCH -J jobname          # Job name, you can change it to whatever you want
#SBATCH -o %N.%j.out        # Standard output will be written here
#SBATCH -e %N.%j.err        # Standard error will be written here
#SBATCH -p compute          # Slurm partition, where you want the job to be queued 
#SBATCH --exclusive         # Request exclusive access to a node (all 28 cores, 128GB of RAM) 
#SBATCH --t=00:10:00          # run for 10 hours

module purge
module add modulename
 
command

Examples of where exclusive access is useful:

• Wiki: OpenMP
• Wiki: Matlab Parallel Computing Toolbox

Parallel batch jobs

Intel MPI parallel batch job

#!/bin/bash
#SBATCH -J jobname          # Job name, you can change it to whatever you want
#SBATCH -n 28               # Number of cores 
#SBATCH -N 4                # Number of nodes 
#SBATCH -o %N.%j.out        # Standard output will be written here
#SBATCH -e %N.%j.err        # Standard error will be written here
#SBATCH -p compute          # Slurm partition, where you want the job to be queued 
#SBATCH --t=00:10:00          # run for 10 hours
 
module purge
module add intel/2017 
 
mpirun command

MVAPICH parallel batch job

#!/bin/bash
#SBATCH -J jobname          # Job name, you can change it to whatever you want
#SBATCH -n 28               # Number of cores 
#SBATCH -N 4                # Number of nodes 
#SBATCH -o %N.%j.out        # Standard output will be written here
#SBATCH -e %N.%j.err        # Standard error will be written here
#SBATCH -p compute          # Slurm partition, where you want the job to be queued 
#SBATCH --t=00:10:00          # run for 10 hours
 
module purge
module add mvapich2/2.2/gcc-6.3.0 
 
mpirun command

There are various options for the mvapich2 module, see mvapich2

OpenMPI parallel batch job

#!/bin/bash
#SBATCH -J jobname          # Job name, you can change it to whatever you want
#SBATCH -n 28               # Number of cores 
#SBATCH -N 4                # Number of nodes 
#SBATCH -o %N.%j.out        # Standard output will be written here
#SBATCH -e %N.%j.err        # Standard error will be written here
#SBATCH -p compute          # Slurm partition, where you want the job to be queued 
#SBATCH --t=00:10:00          # run for 10 hours
 
module purge
module add openmpi/1.10.5/gcc-6.3.0 
 
mpirun command

There are various options for the OpenMPI module, see openmpi

High memory batch job

If you task requires more memory than the standard provision (approximately 4GB) then you need to include a directive in your submission script to request the appropriate resource. The standard compute nodes have 128GB of RAM available and there are dedicated high memory nodes which have a total of 1TB of RAM. If your job requires more than 128GB of RAM, then submit to the highmem partition.

The following job submission script runs on the highmem partition and uses the #SBATCH --mem flag to request 500GB of RAM, which results in three things:

• The job will only be allocated to a node with this much memory available
• No other jobs will be allocated to this node unless their memory requirements fit in with the remaining available memory
• If the job exceeds this requested value, the task will terminate

#!/bin/bash
#SBATCH -J jobname          # Job name, you can change it to whatever you want
#SBATCH -n 1                # Number of cores 
#SBATCH -o %N.%j.out        # Standard output will be written here
#SBATCH -e %N.%j.err        # Standard error will be written here
#SBATCH -p highmem          # Slurm partition, where you want the job to be queued 
#SBATCH --mem=500G
#SBATCH --t=00:10:00          # run for 10 hours
 
module purge
module add modulename
 
command

If a job exceeds the requested about of memory, it will terminate with an error message similar to the following (a job which ran with a memory limit of 2GB):

slurmstepd: Step 307110.0 exceeded memory limit (23933492 > 2097152), being killed
srun: Job step aborted: Waiting up to 32 seconds for job step to finish.
srun: got SIGCONT
slurmstepd: Exceeded job memory limit

GPU batch job

#!/bin/bash
#SBATCH -J jobname          # Job name, you can change it to whatever you want
#SBATCH -n 1                # Number of cores 
#SBATCH -o %N.%j.out        # Standard output will be written here
#SBATCH -e %N.%j.err        # Standard error will be written here
#SBATCH --gres=gpu          # use the GPU resource not the CPU
#SBATCH -p gpu              # Slurm partition, where you want the job to be queued 
 #SBATCH --t=00:00:40          # run for 40 minutes

module purge
module add modulename
 
command

• if you require more than one GPU to be used within a single node use the sbatch command #SBATCH --gres=gpu:N (where N is the number required). In some cases the GPU family name should also be included #SBATCH --gres=gpu:tesla:2 (2 GPUs).

Array batch job

An array batch job allows multiple jobs to be executed with identical parameters based on a single job submission. By using the directive #SBATCH --array 1-10 the same job will be run 10 times. The indexes specification identifies what array index values should be used. Multiple values may be specified using a comma separated list and/or a range of values with a "-" separator. For example, "--array=0-15" or "--array=0,6,16-32".

A step function can also be specified with a suffix containing a colon and number. For example, "--array=0-15:4" is equivalent to "--array=0,4,8,12". A maximum number of simultaneously running tasks from the job array may be specified using a "%" separator. For example "--array=0-15%4" will limit the number of simultaneously running tasks from this job array to 4.

The variable $SLURM_ARRAY_TASK_ID can be used within the batch script, being replaced by the index of the job, for example as part of the input or data filename etc.

When the batch script below is submitted, 10 jobs will run resulting the in command being run with the first argument corresponding to the array element of that task, for instance: command 1, command 2, through to command 10. The output of each of these tasks will be logged to a different out and err file, with the format <node job ran on>.<job ID>.<array index>.out <node job ran on>.<job ID>.<array index>.err

#!/bin/bash
#SBATCH -J jobname          # Job name, you can change it to whatever you want
#SBATCH -n 1                # Number of cores 
#SBATCH -o %N.%A.%a.out     # Standard output will be written here
#SBATCH -e %N.%A.%a.err     # Standard error will be written here
#SBATCH -p compute          # Slurm partition, where you want the job to be queued 
#SBATCH --array 1-10
 
module purge
module add modulename
 
command $SLURM_ARRAY_TASK_ID

Advanced Options

• Job time (to be added)
• Dependencies (to be added)
• Reservations (to be added)

Home Application support General Training Programming support