Turning Fragmented GPU Clusters Into One Elastic Compute Pool
Talkat KubeCon + CloudNativeCon + OpenInfra Summit + PyTorch Conference China 2026 · 2026-09-07..09-09·
summary
Describes building an Elastic Compute Pool that federates fragmented GPU capacity across Kubernetes clusters into a single global scheduling domain to handle bursty AI workloads. Explains extensions to Kubernetes Dynamic Resource Allocation (DRA) and integration with Kueue to add application-level topology awareness, failover-aware placement, and cross-cluster resource modeling, and discusses scheduler/resource-management design choices and operational lessons from running training and inference across hundred-thousand-GPU infrastructure.
AI-written summary · about
related talks
32mSmart GPU Management: Dynamic Pooling, Sharing, and Scheduling for AI Workloads in Kubernetes
30mGPU Sharing at CERN: Cutting the Cake Without Losing a Slice!
22mBeware of Fragmentation: Scheduling GPU-Sharing Workloads with Fragmentation...
32mScaling Kubernetes Clusters for Generative Models: Managing GPU Resources for AI App...
31mAI, CERN, and the Quest for GPU Custody: How CERN Leverages DRA for Efficient GPU Sharing
32mScaling AI Workloads with Kubernetes: Sharing GPU Resources Across Multiple Containers
38mImproving GPU Utilization using Kubernetes
28mGPU Reservations: Maximizing Utilization and Fairness Across Teams