Partitions and QoS

Owned by Josh Burks, created
Last updated: Aug 26, 2024 by William Dizon
4 min read

Overview

Both Sol and Phoenix use the same partition and QoS options. Reference this page to help select the best option for your job.

If not explicitly defined, jobs will default to the HTC partition and public QoS.

Partitions

general

The general-use partition comprises all Research Computing-owned nodes. This partition has a wall time limit of 7 days. CPU-only, GPU-accelerated, and FPGA-accelerated jobs will typically use this.

#SBATCH -p general #SBATCH -t 7-00:00:00 interactive -p general

htc

The high-throughput partition is aimed at jobs that can complete within a four-hour walltime. Jobs that can fit within this window have a scheduling advantage, as this partition includes not only Research Computing-owned nodes, but also privately-owned nodes. As part of the arrangement with infrastructure-purchasing labs, htc jobs can run without risk of pre-emption when using the public qos.

#SBATCH -p htc #SBATCH -t 0-04:00:00 interactive -p htc -t 0-4

highmem

The highmem partition is aimed at jobs that can require an extra amount of memory that cannot be satisfied by the regular compute nodes. The computing power is the same as regular compute nodes, just with greater upper-end memory capacity (up to 2TB). The highmem partition is currently capped to 48 hours (2 days) walltime for any given job. If longer than two days is required, a lengthier QOS is available to extend that to 7 days. To use this QOS, we invite you to reach out to the Research Computing staff, where we can help evaluate your job and its fitness to be on these nodes.

#SBATCH -p highmem #SBATCH -t 2-00:00:00 interactive -p highmem --mem=1000G

lightwork

The lightwork partition is aimed at jobs that require relatively less computing power than typical supercomputing jobs and may stay idle for larger amounts of time. Great examples of this would be creating mamba environments, compiling software, VSCode tunnels, or basic software tests. The aux-interactive command will automatically allocate on the lightwork partition

Jobs that utilize cores to their full potential are more appropriately used in htc or general partitions, where cores are not shared/oversubscribed. Jobs that use full cores to > 99% for a continued duration of time, or jobs that request excessive resources prohibiting other users from using this partition, are subject to cancellation. Repeated misuse of this partition will result in ineligibility from using lightwork going forward.

QOS

public

The public QOS should be used for any jobs that use public Research Computing Resources. This includes any job submitted to the above-listed partitions. Under most circumstances, public is the preferred QOS.

debug

debug is a special QOS for testing your sbatch jobs for syntax errors and pre-compute testing. When setting up your workflows, rather than submit to the normal compute partitions, submitting to debug offers the following advantages:

  • Provides a much quicker turnaround in troubleshooting your scripts.
    Are there syntax errors as your script progresses?
    Are all paths and modules properly set to ensure a successful run?

  • Has shorter expected times for jobs to start

  • Using the debug partition with a limited (small) dataset means you can quickly confirm the validity of your pipelines: you can then switch to your full dataset with greater confidence it will complete with the desired output.

The debug QOS works with the general and htc partitions for walltimes up to 15 minutes.

private

The private QOS signifies a willingness to have your job flagged as pre-emptable (cancellable by hardware owners) as a trade-off for being able to use that privately owned resource longer than the protected four-hour htc partition offering. This can sometimes greatly reduce the time-to-start for jobs, but is heavily dependent on the aggregate supercomputer load. Pre-emptable jobs will be canceled if and only if the private owners submit a job to those resources.

Examples of this might be to use privately owned GPUs, especially where checkpointing is used.

Job preemption will occur in either of the following circumstances:

  • a member of the hardware-owning group schedules a job that cannot be satisfied because of the private job allocation, e.g., grp_labname.

grp_labname

Any resources purchased by independent labs will be given a grp_labname QOS. For the users of the group, there is no fairshare impact for using owned hardware.

Note, that the general partition is still used by hardware owners, but the QOS itself limits the job to the group’s eligible hardware. grp_labname jobs will not preempt public jobs 4 hours or shorter, but will preempt private jobs of any length.

long

The long QOS permits runtimes on Research Computing-owned hardware to exceed the 7-day maximum and extends this to 14 days. It should not be used with interactive, only SBATCH

This is a special-case QOS not available by default, it is granted on a user-by-user basis; if you are interested in using this QOS, please be ready to share a job ID demonstrating the need and effective use of existing core allocations. If you have any questions, feel free to ask staff and also explore the Slurm EFFiciency reporter:

 

Additional Help

Arizona State University