Context
From the May 28 demo meeting, Thameem raised concern that fragmented single-GPU reservations could prevent teams from accessing full clusters for larger workloads. The group discussed implementing packing logic to guide users toward partially-utilized clusters.
Current State
- No allocation or bin-packing algorithm exists
- Users manually select a cluster from a dropdown when creating a reservation
- There is no guidance or recommendation on which cluster to use
Proposed Changes
- Implement a packing/bin-packing algorithm that recommends clusters with existing partial reservations before allocating fresh clusters
- When a user requests N GPUs, suggest the cluster that best fits the request while minimizing fragmentation
- Surface packing recommendations in the reservation form (e.g., "Cluster X has 4/8 GPUs available — consider using it")
- Optionally enforce packing via the auto-select feature rather than just recommending
Dependencies
- Requires partial GPU reservations to be implemented first
Context
From the May 28 demo meeting, Thameem raised concern that fragmented single-GPU reservations could prevent teams from accessing full clusters for larger workloads. The group discussed implementing packing logic to guide users toward partially-utilized clusters.
Current State
Proposed Changes
Dependencies