Resource as You Wish: Collaborative Reservation and Allocation b...
summary
Shows a working system that uses a control-plane operator, a custom scheduler, extended resource names and a node device plugin to enforce PCIe-topology aware allocations of AI accelerators (XPUs), including a getPreferredAllocation hint flow so specific cards behind the same PCIe switch get chosen. Custom resources expose topology and allocation state; the solution runs today and will appear in the Red Hat OpenShift catalog, while Dynamic Resource Allocation (DRA) is noted as a more powerful but still alpha alternative.
AI-written summary · about
chapters (7)
- 0:00AI accelerator challenges
- 2:53Kubernetes device plugins
- 7:10Device allocation in Kubernetes
- 13:13AI accelerator allocation challenges
- 19:25Custom scheduler solution
- 21:57Comparing resource allocation
- 23:00Summary and future
related talks
50mDynamic Resource Allocation - the path towards GA
29mWhat Can I Get You? An Introduction to Dynamic Resource Allocation
27mUncharted Waters: Dynamic Resource Allocation for Networking
05mResource Roulette: Winning the Kubernetes Allocation Game
35mResource Management WG Deep Dive
26mResource Orchestration of HPC on Kubernetes: Where We Are Now and the Journey Ahead!
34mKubernetes DRA Architecture: Scheduling, Status, and Topology at Scale
31mMore Nodes, More Problems: Solving Multi-Host GPU/TPU Scheduling With Dynamic Resource Allocation