[Ofmfwg] Sunfish meeting notes for February 2, 2026

[Lee, Peter]

  *   Near Node Flash Use Case
     *   Continue working on the flow diagram to include the AI scheduling support.
     *   Looking to support to following architecture.
        *   Sunfish is responsible for discovering and aggregating raw NVMe resources across different NVMe fabrics. Sunfish acts as an accountant and inventory system, not as a volume and workload planner. Sunfish only tracks what NVMe endpoints exist and publishes changes via events when there are changes to the namespaces.
        *   Swordfish resource manager queries Sunfish to find available NVMe endpoints and their properties. Swordfish resource manager is the entity that issues commands directly to the NVMe resources to make any changes to the storage resources. Swordfish resource manager maintains all the high-level volume information and can offload this volume inventory to Sunfish.
        *   Knapsack scheduler works with Swordfish resource manager to configure the required storage resources needed for the different workloads or job requests. Knapsack scheduler also interacts with Flux broker 0 to get resource requirements for workloads or job requests and pass the resultant configured storage resources for the job requests to Flux broker 0 for it to hand out to the different compute hosts.
     *   Mike is working on a model that takes the resources from Sunfish and the requirements from job requests to schedule the resources. When you run inference on this model, the result is an approximation. Some result may be better than others. Still working on the model.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openfabrics.org/pipermail/ofmfwg/attachments/20260202/da08d22f/attachment.htm>