Queue Policies

7 minute read

Start Times

Start times are a great way to track your jobs:

squeue -u $USER --start

Start times are rough estimates based on the current state of the queue.

Partition Quotas

Each partition has a specific usecase. Below outlines each partition, it’s usecase, as well as any job/user/group limits that are in place. Empty boxes imply no limit, but is still limited by the next higher limit. Job limits are capped by user limits, and user limits are capped by group limits.

In addition to the above limits, there is:

• A 768 core group limit that spans across all users in a group across all partitions.
• A 8 GPU group limit that spans across all users in a group across all GPU-enabled partitions.

Attempting to allocate more member than a node can support, eg 500GB on an Intel node, will cause the job to immediately fail. Limits are for actively running jobs, and any newly queued job that exceeds a limit will be queued until resources become available. If you require additional resourced beyond the listed limits, please see the “Additional Resource Request” section below.

Partition quotas can also be viewed on the cluster using the slurm_limits command.

Additionally, users can have up to 5000 jobs in queue/running at the same time. Attempting to queue more than 5000 jobs will cause jobs submissions to fail with the reason “MaxSubmitJobLimit”.

External Labs

Labs external to UCR will have reduced resource limits as follows:

• Labs will have a CPU quota of 256 cores across all lab users across all partitions
• Per user CPU quotas on epyc, intel, batch, and short will be 128 cores
• Per user CPU quotas on highmem will be 16
• GPU quotas on the gpu partition will be 4 per-lab, and 2 per-user
• CPU quotas on the gpu partition will be 24 per-user and 8 per-job

Private Node Ownership

Labs have the ability to purchase nodes and connect them to the cluster for increased quotas. More information can be found in the Ownership Model section of our Access page.

Additional Resource Request

Sometimes, whether it be due to deadlines or technical limitations, more resources might be needed than are supplied by default. If you require a temporary increase in quotas, please reach out to support@hpcc.ucr.edu with a short justification as to why additional resources are needed. The following are typical circumstances that could justify increased quotas:

• Urgent Deadlines: ie. Grant submissions, conference presentations, paper deadlines
• Special Technical Needs: The limits do not meet the technical requirements for the program(s) that are trying to be ran.

The amount of additional resources, the length of time that the resources are needed, and the frequency of the requests are all factors that determine whether your request will be accepted. It also must be within the capacity of the HPCC’s infrastructure while also ensuring minimal disruption to other users. The final decision of approving exception requests, and how many extra resources to provide, will be decided by the HPCC Staff, the Director, and in exceptional cases the HPCC Oversight Committee.

Requests limited by unoptimized code/datasets or strictly for the sake of convenience will be denied.

Example Scenarios

Per-Job Limit

A job is submitted on the gpu partition. The job requests 32 cores.

This job will not be able to be submitted, as 32 cores is above the partition’s 16 core per-job limit.

Per-User Limit

You submit a job the highmem partition, requesting 32 cores.

This job will start successfully, as it is within the partition’s core limit.

You submit a second job while the first job is still running. The new job is requesting 32 cores.

Because you are at your per-user core limit on the highmem partition, the second job will be queued until the first job finishes.

Per-Lab Limit

User A submits a job requesting 384 cores. User B submits a job requesting 384 cores.

Because each user is within their per-user limits and the lab is within their limit, the jobs will run in parallel.

User C submits a job, requesting 16 cores.

Because User A and User B are using all 768 cores within the lab, User C’s job will be queued until either User A’s or User B’s jobs finishes.

Changing Partitions

In srun commands and sbatch scripts, the -p or --partition flag controls which partition/queue a job will run on. For example, using -p epyc will have your job queued and ran on the epyc partition. For more examples and information on running jobs, see the Managing Jobs page of our documentation.

Fair-Share

Users that have not submitted any jobs in a long time usually have a higher priority over others that have ran jobs recently. Thus the estimated start times can be extended to allow everyone their fair share of the system. This prevents a few large groups from dominating the queuing system for long periods of time.

You can see with the sqmore command what priority your job has (list is sorted from lowest to highest priority). You can also check to see how your group’s priority is compared to other groups on the cluster with the “sshare” command.

For example:

sshare

It may also be useful to see your entire group’s fairshare score and who has used the most shares:

sshare -A $GROUP --all

Lastley, if you only want to see your own fairshare score:

sshare -A $GROUP -u $USER

The fairshare score is a number between 0 and 1. The best score being 1, and the worst being 0. The fairshare score approches zero the more resource you (or your group) consume. Your individual consumption of resources (usage) does affect your entire group’s fiarshare score. The affects of your running/completed jobs on your fairshare score are halved each day (half-life). Thus, after waiting several days without running any jobs, you should see an improvment in your fairshare score.

Here is a very good explaination of fairshare.

Priority

The fairshare score and jobs queue wait time is used to calculate your job’s priority. You can use the sprio command to check the priority of your jobs:

sprio -u $USER

Even if your group has a lower fairshare score, your job may still have a very high priority. This would be likely due to the job’s queue wait time, and it should start as soon as possible regardless of fairshare score. You can use the sqmore command to see a list of all jobs sorted by priority.

Backfill

Some small jobs may start before yours, only if they can complete before yours starts and thus not negatively affecting your start time.

Priority Partition

Some groups on our system have purchased additional hardware. These nodes will not be affected by the fairshare score. This is because jobs submitted to the group’s partition will be evaluated first before any other jobs that have been submitted to those nodes from a different partition.