On the Mill there are multiple partitions that you can submit jobs to. The optimal partition will depend on your resource needs. A summary of the various partitions is below, with more details on the individual partitions following.
| |
Maximum Time Limit |
Description |
| priority partitions |
28 days |
For investors. |
| requeue |
1 day |
Any user may request idle capacity for their jobs. Jobs in the requeue partition may be interrupted at any time for paid users. |
| std-cpu |
14 days |
Standard pay-as-you go queue for CPU jobs. Use requires a MOU and an account for billing. |
| std-v100 |
14 days |
Standard pay-as-you go queue for GPU jobs on V100 hardware. Use requires a MOU and an account for billing. |
| std-h100 |
14 days |
Standard pay-as-you go queue for GPU jobs on H100 hardware. Use requires a MOU and an account for billing. |
| interactive |
4 hours |
For short interactive testing, interactive debugging, and general interactive jobs. Use this for light testing as opposed to the login node. Max of 8 CPUs per job. No GPUs are available. Interactive jobs will not be requeued. |
| class |
4 hours |
This partition is for students working on coursework. |
Priority Partitions
Up to 50% of the total compute capacity in the Mill can be leased by individual research groups. If your lab has purchased or leased computing resources in the Mill, then you will have access to a lab partition (commonly named as your PIs's last name). Submitting your jobs to this partition will generally offer the best performance.
Requeue Partition
The requeue partition is the default partition. This is where jobs will be queued if no partition is set in the SLURM submission script. When nodes are idle they will accept jobs from the requeue partition at no cost. These jobs will be interrupted and sent back to the queue should a paid submit a job, and the paid job will start immediately. This makes requeue an ideal partition for short jobs that are unlikely to be interrupted before they finish and jobs with robust checkpointing that can make non-continuous progress in the case of an interruption.
Standard Partitions
The standard partitions are: std-cpu, std-v100, and std-h100. These partitions are pay-as-you go and have priority over requeue jobs. If a job is submitted to a standard partition, it can send a running requeue job back into the queue to start immediately.
Using the standard queues requires a MOU and an account set up for billing. A lab's account must be specified with --account=<name> and the submitting user must be a member of that lab group for the job to be accepted to the queue.
The std-v100 and std-h100 partitions have GPU hardware available. Due to CUDA support differences between the two cards, please use module load cuda-toolkit/12.9 for use on the V100s and module load cuda-toolkit/13.3 for the H100s.
Interactive
The interactive partition is intended for interactive sessions rather than asynchronous jobs. Good use cases include things like graphical output for visualizations, rapid prototyping, and debugging. If you need graphical output, see the instructions on enabling X-forwarding.
Class
The class partition is used for student coursework. This is the only partition that students will have access to.