Question: how will Grove support intra-node NUMA-aware scheduling?

[yankay]

Hi maintainers, thanks for GREP-244.

I noticed TopologyDomainNuma = "numa" is defined in the API and Story 3 (NUMA-Aware GPU Benchmarking) is listed as a motivating user story, but I cannot find any controller/scheduler logic that consumes it.

The current ClusterTopologyBinding model requires each domain to map to a Node label key. That works for rack/zone/host (one Node → one label value), but NUMA sockets live inside a Node and are never exposed as Node labels. So Story 3 ("2 GPUs from an 8-GPU node on the same NUMA node") does not seem reachable through node-label pack alone.

Related context:

• ai-dynamo/dynamo#10171 — NUMA-aware placement discussion.
• KEP-4381 Partial GPU Allocation.
• kubernetes/kubernetes#132296 — standardized resource.kubernetes.io/pcieRoot DRA attribute for intra-node GPU↔NIC alignment.

Question: which direction is planned?

1. DRA path — rely on upstream DRA (pcieRoot/NUMA attributes) ; Grove stays at the node-domain layer and possibly injects ResourceClaim selectors (similar to auto-MNNVL).
2. Scheduler path — push intra-node NUMA awareness into the scheduler backend (e.g. KAI, see kai-scheduler/KAI-Scheduler#1598,Volcano numa-aware).
3. Others

Happy to help with a prototype or follow-up GREP once the direction is clear. Thanks!