GPU Sharing at CERN: Cutting the Cake Without Losing a Slice!
Talkat KubeCon + CloudNativeCon Europe 2025 · on Thursday, 2025-04-03· at 16:45· Level 1 | Hall Entrance S10 | Room D·
summary
Centralize GPUs with Kubernetes DRA 1.42 beta and an NVIDIA DRA driver to allocate time-slice, MPS or MIG without node relabels; benchmarks on an A1 40GB card show time slicing is easiest to enable but adds context-switch overhead and extra reserved memory that can cause OOMs, MPS runs parallel workloads well when GPUs are underutilized, and MIG gives isolation with non‑linear scaling — LQFT gains about 3× speedup moving 2G/10GB→4G/20GB, while MLPF prefers smaller partitions; dynamic MIG lands in Kubernetes 1.43.
AI-written summary · about
chapters (18)
- 0:00Introduction and CERN context
- 1:27Why GPUs at CERN
- 2:29Centralizing GPU resources
- 3:55GPU sharing options
- 5:42Time slicing vs MPS vs MIG
- 8:55Manual GPU partitioning challenges
- 10:40Dynamic Resource Allocation (DRA)
- 13:09DRA in action demo
- 15:38DRA benefits and system overview
- 16:23GPU sharing limitations
- 17:06Benchmarking setup
- 18:30Use cases: LQFT and MLPF
- 19:14Time slicing performance results
- 20:59MPS performance results
- 22:24Workload-aware GPU sharing
- 23:26MIG partitioning insights
- 25:13MIG isolation advantages
- 25:55Final conclusions and Q&A
related talks
31mAI, CERN, and the Quest for GPU Custody: How CERN Leverages DRA for Efficient GPU Sharing
33mAI workloads on Kubernetes - How to maximize GPU utilization and cut costs
38mImproving GPU Utilization using Kubernetes
32mScaling AI Workloads with Kubernetes: Sharing GPU Resources Across Multiple Containers
32mScaling Kubernetes Clusters for Generative Models: Managing GPU Resources for AI App...- 32mScaling AI Workloads with Kubernetes: Sharing GPU Resources Across Multiple Containers
34mDeclarative Device Virtualization: Orchestrating GPUs & Hardware in Cloud Native Environments
27mIncremental GPU Slicing in Action