AI, CERN, and the Quest for GPU Custody: How CERN Leverages DRA for Efficient GPU Sharing
Talkat Cloud Native + Kubernetes AI Day Europe 2025 · on Tuesday, 2025-04-01· at 15:55· Level 1 | Hall Entrance N10 | Room G·
summary
CERN centralizes GPUs into a shared pool and reduces idle time by combining time-slicing, NVIDIA MPS, and MIG (up to seven isolated instances) with Dynamic Resource Allocation (DRA) to partition and provision GPUs on demand. DRA removes per-node prelabeling pain for nodes with 2/4/8 GPUs, lets users request specific device configs and groups via resource claims, and provides a compute domain primitive that sets up ephemeral NVLink/IMAX channels for cross-node GPU memory sharing — demonstrated working. Kubernetes 1.32 exposes the feature gate; 1.33 adds dynamic MIG provisioning; GA aims later this year.
AI-written summary · about
chapters (14)
- 0:00CERN and GPU challenges
- 2:09Centralizing GPUs for efficiency
- 3:30GPU sharing options and limits
- 7:23Manual GPU partitioning pain
- 9:10DRA for dynamic GPU allocation
- 9:52What is Dynamic Resource Allocation
- 13:03Limitations of device plugins
- 15:54DRA example: Controlled GPU sharing
- 17:41DRA example: Advanced device requests
- 18:44DRA example: Dynamic MIG provisioning
- 19:49DRA example: Multi-node GPU comms
- 23:22Live demo: Compute domain
- 25:12Future of DRA and features
- 25:56Q&A
related talks
30mGPU Sharing at CERN: Cutting the Cake Without Losing a Slice!
34mDeclarative Device Virtualization: Orchestrating GPUs & Hardware in Cloud Native Environments
30mRunning AI Workloads in Containers and Kubernetes
32mLeveraging DRA for Conquering ML Workload Challenges on...
28mGPU Reservations: Maximizing Utilization and Fairness Across Teams
38mImproving GPU Utilization using Kubernetes
32mScaling AI Workloads with Kubernetes: Sharing GPU Resources Across Multiple Containers
33mAI workloads on Kubernetes - How to maximize GPU utilization and cut costs