Automatic detection of underutilized jobs

We are working on a system that sends users alerts when their jobs are underutilizing or overloading the compute nodes. We are going to do a multi-stage release, with the first stage starting this week.

Users with jobs that are severly underutilizing resources—including requesting a GPU and not using it, submitting an interactive job and not using it, and using very low CPU count on a fully allocated compute node—will receive e-mail alerts describing the problem, suggesting ways to solve it, and providing a link to details about the job.

Please pay attention to these emails and try to rectify the problem. This will help other users who may be waiting for the resources used by inefficient jobs to become available, and it could help you find ways to improve the performance of your jobs.

As this is the first step in a more complete implementation, we are interested in feedback, especially if the suggestions in the email are not appropriate for your flagged job. You can submit feedback by emailing [email protected].