Dynamic Resource Allocation - the path towards GA
summary
Dynamic Resource Allocation (DRA) is an alpha API from 1.26 that now tests two approaches: opaque parameters (vendor CRDs, custom pod scheduling context, delayed allocation) and structured parameters (built‑in resource claim parameters, resource slices, scheduler‑side matching). An example Helm‑installed driver running on kind demonstrates live allocation and sharing of GPUs (simulates real GPUs, kubelet plugin + controller, NodeAllocationState CRD) and already covers 6 of 12 NVIDIA GPU use cases; goal is 10/12 by 1.31, missing MIG/partitioning and management pods.
AI-written summary · about
chapters (8)
- 0:00Introduction to DRA
- 1:02Opaque parameters approach
- 7:02Example driver demonstration
- 17:43Structured parameters design
- 22:48Named resources model
- 30:58Use cases and roadmap
- 35:35Future cleanup and timeline
- 43:24Q&A and closing remarks
related talks
29mWhat Can I Get You? An Introduction to Dynamic Resource Allocation
42mGuide Us on Next Steps in Kubernetes Device Manage...
27mUncharted Waters: Dynamic Resource Allocation for Networking
24mFit-to-Serve: How a New DRA Capability for Dynamic Device...
34mKubernetes DRA Architecture: Scheduling, Status, and Topology at Scale
26mResource as You Wish: Collaborative Reservation and Allocation b...
32mLeveraging DRA for Conquering ML Workload Challenges on...
30mKubernetes WG Device Management - GPUs, TPUs, NICs and More With DRA