Choosing a Partition

This page is a guide for how to choose a partition for your HPC job.  For a quick reference of available partitions, refer to our Partition List.

Introduction

Each user on the HPC system has access to certain partitions (a.ka. "queues") on the system.  Each partition provides access to different nodes, wait times, and maximum job runtimes.  Some partitions are general access, which means every HPC user has access to them: genacc_q, backfill, quicktest, condor and backfill2.  There are also partitions which only certain user groups have access to.  These are called owner-based partitions.  In order to determine which partitions you have access to, type the following command into the terminal once you are logged into the HPC system:

rcctool my:partitions

The output of this command will display a list of which partitions that you have access to.  From here, you must decide which partition you want to submit your job to.

The Basics of Selecting a Partition

  • If you have an owner-based partition, you should generally use that. Your job will likely start sooner and have access to more resources than if you use general access.  However, there are many cases in which it is better to run your job in a general access partition, even if you have access to an owner-based partition.
  • If you do not have access to an owner-based partition, you should use one of the general access partitions.

General Access Partitions

  • genacc_q: General Access Partition
  • backfill: The Backfill Partition
  • backfill2: Secondary Backfill Partition
  • quicktest: Testing Partition for Short-Runtime Jobs
  • condor: Very long-Runtime Jobs

Pros and Cons of General Access Partitions

General Access Main Partition (HPC Name: genacc_q)

The genacc_q partition is intended for long-running jobs requiring multiple nodes and shared memory

  1. General Access is ideal for jobs that use MPI and need massive scalability.
  2. The partition is also ideal for multi-node shared memory jobs such as those that use a mixture of OpenMP and MPI that will take longer than few hours to complete.
  3. Each job is limited to 14 days of runtime.

Backfill Partition (HPC Name: backfill)

The backfill partition is intended for jobs that only require short run-times.

  1. Backfill is meant to run jobs which only require less than 4 hours of runtime.
  2. Jobs on backfill tend to start fairly soon after the job is sumbitted.
  3. Jobs running on Backfill have access to newer nodes than genacc_q.

Secondary Backfill Partition (HPC Name: backfill2)

The backfill2 partition is very similar to the main backfill parititon, but jobs can be pre-empted:

  1. Jobs are also limited to 4 hours of runtime, just as in the main backfill partition.
  2. Using backfill2 grants your job access to all available nodes on HPC.
  3. However, this includes owner-based nodes, and backfill2 jobs run the risk of being pre-empted and cancelled automatically by owner submitted jobs.
  4. Jobs start very soon after submission.
  5. This is ideal if you want a job started quickly and are okay with the potential of having to re-submit your job if it is pre-empted by an owner-based job.

Condor (HPC Name: condor)

The condor partition is intended to be used for jobs which will require very long runtimes where each task needs only a single compute node with multiple cores.

  1. Jobs can run for up to 90 days
  2. Ideal for single-node jobs (can run on 16 cores or less), shared-nothing jobs where OpenMP, MPI, and other message-passing is not needed (e.g. map-reduce, etc)
  3. Consists of older hardware

Quick Test Partition (HPC Name: quicktest)

The quicktest partition is mainly intended for prototyping and testing of code and scripts.

  1. Jobs are limited to 10 minutes of runtime.
  2. Jobs are also limited to a single node.
  3. This is ideal for quickly testing your workflow and software programs on our scheduler.

Need GPUs?

If your job requires GPUs, there are two options available to you:

  1. use the GPU nodes available in backfill2; or,
  2. purchase GPU resources for an owner-based partition.

For further details, see GPU Access Documentation.