DRA in KubeVirt: Overcoming Challenges & Implementing Changes
summary
Proposal replaces Kubernetes Device Plugin with DRA integration into KubeVirt to solve coordination, deletion, and cluster-state races encountered running GeForce Now Cloud VMs with pass-through GPU and 2:1 and 4:1 vGPU classes. Uses Kubernetes 1.30/1.32 node allocation state objects, ResourceClaims, and generates CDI JSON on allocate; two handoff methods are shown (env vars or a plugin CDI JSON sidecar that converts allocation metadata into libvirt domain XML). Work is a prototype design aligned with K8s alpha/beta DRA lifecycle.
AI-written summary · about
chapters (12)
- 0:00DRA in KubeVirt intro
- 1:26GPU use cases GeForce Now
- 5:28Device Plugin limits
- 9:30Dynamic Resource Allocation intro
- 12:26DRA solves DP issues
- 13:53KubeVirt + DRA integration
- 14:14DRA API changes
- 16:21Allocating devices to launcher
- 20:13Passing devices to KVM
- 22:20Extensible sidecar device handling
- 25:38DRA + KubeVirt future
- 26:29Q&A
related talks
37mKubeVirt Deep Dive: Virtualized GPU Workloads on KubeVirt
30mKubernetes WG Device Management - GPUs, TPUs, NICs and More With DRA
29mWhat Can I Get You? An Introduction to Dynamic Resource Allocation
42mGuide Us on Next Steps in Kubernetes Device Manage...
36mIt’s All for the Users. More Durable, Secure, and Pluggable. KubeVirt v0.53
32mLeveraging DRA for Conquering ML Workload Challenges on...
34mKubernetes DRA Architecture: Scheduling, Status, and Topology at Scale
50mDynamic Resource Allocation - the path towards GA