Skip to main content

QoS (Quality of Service)

warning

These are our old docs! Please see our new docs by clicking on the M3 docs dropdown and selecting "New M3 docs".

These old docs were roughly converted from our old format. As a result, this copy is not identical to our previous docs. You can still find our old docs at https://old-docs.massive.org.au/.

You may notice some formatting and structural issues with these old docs. We will not resolve these, this is here purely for backwards compatibility to ensure old URLs do not die.

We have implemented Quality of Service (QoS) starting 6th of June 2018.

The QoS can be added to each job that is submitted to Slurm. The quality of service associated with a job will affect the job in three ways:

  1. Job Scheduling Priority
  2. Job Preemption
  3. Job Limits

How to run jobs with QoS

A table to show the differences between QoS:

QueueDescriptionMax WalltimeMax GPU per userMax CPU per userPriorityPreemption
normalDefault QoS for every job7 days420050No
rtqQoS for interactive job48 hours672200No
irqQoS for interruptable job7 daysNo limitNo limit200Yes
shortqQoS for job with short walltime30 mins10280250No

Explanation

These QoS are applied to the partition comp.

--qos=normal

This is the QoS for all the jobs that do not specify a QoS. Jobs that run here won't be interrupted.

--qos=rtq

This is intended to be used by jobs that have an instrument or a real-time scenario and therefore can't be interrupted and must be available on demand. You can only use a few CPUs and GPUs, but jobs will start as soon as possible (before normal jobs).

--qos=irq This is intended to be used by jobs that are interruptible. To use the irq you have to be prepared to either restart from scratch (if the job was short anyway) or restart from a checkpoint. Jobs will start very quickly and use all the available resources but may be stopped at short notice to allow shortq or rtq jobs to run.

The mechanism to checkpoint depends on the software that you are running.

Please contact us if you have any questions with regards to job checkpointing.

--qos=shortq This is intended to be used by short and uninterruptible jobs. These jobs will run before normal but the walltime is limited.

An example of Slurm GPU job script

#!/bin/bash
#SBATCH --job-name=MyJob
#SBATCH --account=<my_account>
#SBATCH --qos=<irq,shortq,rtq>
#SBATCH --gres=gpu:V100:1
#SBATCH --ntasks=2

<GPU processing program>

An example Slurm job script

#!/bin/bash
#SBATCH --job-name=MyJob
#SBATCH --account=<my_account>
#SBATCH --qos=<irq,shortq,rtq>
#SBATCH --ntasks=2

openmpi/1.10.7-mlx
mpirun <program>

An example of Slurm job script for Genomics partition

#!/bin/bash
#SBATCH --job-name=MyJob
#SBATCH --account=<my_account>
#SBATCH --qos=genomics
#SBATCH --partition=genomics
#SBATCH --ntasks=2

module load bwa