Rules - Read these first
- MDST's flux allocation may only be used by current students working on MDST projects, no exceptions.
- Our Flux allocation may not be used for academic research projects.
- Our Flux allocation may not be used for personal projects not related to MDST.
- Our Flux allocation may not be used for any commercial purposes.
Getting Access to Flux
Before you start, you must be signed up for MDST.
- Get DUO & Make a Flux account
- Ask @stroud (on Slack) to add you to the MDST flux allocation. If you want GPU access, mention this.
- Wait for @stroud to confirm. It can take up to a business day for ARC-TS to process your request.
- Log in to flux and check that this directory exists: /scratch/mdatascienceteam_flux/$USER
If you have never used Flux before, we recommend you check out this tutorial on Github.
If you have used Flux before, note that our mdatascienceteam_flux allocation operates on the preempt queue. This is the only change you will need to make to your PBS scripts when using MDST resources.
Flux is the HPC cluster on campus, operated by ARC-TS. We have two flux allocations for MDST:
- CPUs: 1000
- Memory: 4TB
- Queue: preempt
- GPUs: 2
- CPUs: 4
- Memory: 16GB
- Queue: fluxg
GPU Usage Guidelines
- Interactive jobs are for debugging only
- Limit your interactive jobs to 2 hours
- Do not request more than 2 CPUs + 8 GB mem per GPU. Requesting more than this will potentially prevent jobs from starting on the other GPU.
- Parallelize jobs whenever possible.
- Checkpoint your code at regular intervals by saving models, predictions, etc. This way, if your job crashes, you can start where you left off rather than starting over.