Change to interactive GPU policies

[chtc-users@xxxxxxxxxxx]

Greetings CHTC users, 

This email is for users of the GPUs in CHTC's high throughput computing (HTC) system. 

We are changing our interactive GPU policies to ensure fairness on CHTC’s shared GPUs.

What’s changing:

Starting April 22, condor_ssh_to_job will be unavailable for jobs running on CHTC’s shared GPU machines.

What’s staying the same:

If your group has owned or prioritized access to GPUs, you will still be able to connect to running jobs on said GPUs using condor_ssh_to_job. 

All users will still be able to submit interactive GPU jobs using the condor_submit -i command.

Why we are making the change:

Some users have been using the condor_ssh_to_job functionality to submit long-running sleep jobs, then connect periodically to run interactive commands while leaving the GPU unused for long periods of time. This usage pattern by others limits YOUR throughput, particularly for high-end GPUs, so we are removing this functionality from our shared use GPUs. 

What you can do:

• You can run interactive jobs on general use GPUs. We have expanded our documentation on how to access GPUs this way, and what is currently available.

• Check-in on the output of your jobs while they are running by using condor_tail. See our documentation. 

Your feedback:

If you have questions or comments specifically about this change in policy, please fill out our feedback form.

Sincerely,

Your CHTC Facilitation Team