Kubernetes WG Device Management - GPUs, TPUs, NICs and More With DRA
Talkat KubeCon + CloudNativeCon Europe 2025 · on Thursday, 2025-04-03· at 15:00· Level 3 | ICC Capital Suite 10-12·
summary
DRA reaches beta in Kubernetes 1.32 and the WG moves features toward GA (targeting 1.34); scheduler-side logic, kubelet drivers and the pod resources API are implemented; GPU Operator and NVML back NVIDIA’s DRA driver and an example driver ships as deployable images. 1.33 promotes driver-owned resource claim status to beta, adds alpha partitionable devices, device taints/tolerations, prioritized alternatives, admin-mode namespace gating, and a seamless driver rolling update for daemonsets. NVIDIA demos multi-node NVLink on four nodes (16 GPUs) and reports full-mesh throughput working end-to-end.
AI-written summary · about
chapters (15)
- 0:00Kubernetes Device Management WG intro
- 1:04Device Management WG mission
- 3:08Dynamic Resource Allocation concepts
- 5:58DRA beta in Kubernetes 1.32
- 7:03DRA new features in 1.33
- 9:37DRA GA progress in 1.34
- 11:26DRA future features
- 13:54NVIDIA DRA multi-node GPU
- 18:55NVIDIA DRA demo hardware
- 23:11Q&A backward compatibility
- 24:56Q&A Intel DRA driver
- 25:16Q&A device plugins future
- 26:40Q&A GPU fractionalization
- 27:42Q&A monitoring devices
- 29:49Q&A wrap-up
related talks
42mGuide Us on Next Steps in Kubernetes Device Manage...
35mResource Management WG Deep Dive
29mWhat Can I Get You? An Introduction to Dynamic Resource Allocation
32mLeveraging DRA for Conquering ML Workload Challenges on...
26mState of the GPU(nion)
50mDynamic Resource Allocation - the path towards GA
24mFit-to-Serve: How a New DRA Capability for Dynamic Device...
30mDRA in KubeVirt: Overcoming Challenges & Implementing Changes