Fair-Share
How the scheduler manages fair share of resources amongst all users
To prevent a single user from using all available resources on the cluster we have a range of 'Quality of Service' rules that are enforced. The current rules came into force on the 14/5/2025 and are detailed below:
| Partition | QOS Limits across partition |
| short |
|
| long |
|
| gpu_short |
|
| gpu_long |
|
| Interactive Jobs |
|
At the discretion of IT staff some of these rules can be relaxed on a per-job basis, but where this is done, these jobs will be restricted to one at a time.
As the CPU only batch nodes all service the `short` and `long` partitions, there is an overall limit on the number of threads available across each partition. This is currently set to the following:
- Short: 72 threads per device (288 threads total)
- Long: 36 threads per device (144 threads total)
Your job's position in the queue depends on several factors:
- How many jobs you have run in the last three days
- The CPU threads, memory and GPUs you have requested, in the ratio 1:2:1
- How long you have been waiting for
Interactive jobs have a different usage history to queued jobs, and so submitting lots of jobs or running interactive sessions for long periods/regularly will not effect your priority for jobs of the alternate type.