A Run:ai project uses free GPUs above its quota. What can happen when the owner needs them?
Why A is the answer
The over-quota work can be preempted
Over-quota use borrows idle capacity.
In detail
- Over-quota use borrows idle capacity, and it can be preempted.
- Save checkpoints in over-quota work.
Use case
An illustrative scenario, not a real incident.
A project borrows idle GPUs, and that work is preempted when the owner returns.
See it in different ways
flowchart TD A["Departments"] --> B["Projects with quotas"] B --> C["Over quota when GPUs are free"] C --> D["Preempted when the owner needs them"]
mindmap
root((Sharing))
Run:ai
Project quotas
Over quota, then preempted
Slurm
Fairshare
MIG
GPU instances
Hardware isolationstateDiagram-v2 state "Idle GPUs" as s0 state "Borrowed" as s1 state "Preempted" as s2 [*] --> s0 s0 --> s1: over quota s1 --> s2: owner needs them s2 --> [*]
Key point: Over-quota work can be preempted.
Why the other options are wrong
- B. The cluster stops scheduling
- Scheduling continues.
- C. The owner must wait for a week
- The owner gets GPUs back.
- D. The GPUs are removed for good
- GPUs are not removed.