Smart GPU Management: Dynamic Pooling, Sharing, and Scheduling for AI Workloads in Kubernetes
Talkat KubeCon + CloudNativeCon China 2025 · on Tuesday, 2025-06-10· at 15:30· Level 19 | Crystal Court I·
summary
Covers combining GPU sharing technologies—vCUDA, MPS, and NVIDIA MIG—into a unified Kubernetes workflow that hides GPU types and sharing methods from users. Describes an approach that lets users request memory and core counts; the system auto-selects nodes and the best sharing method, translates requests into optimal GPU partitioning profiles, and dynamically partitions GPUs, using Volcano and HAMi to improve pooling and scheduling for AI workloads.
AI-written summary · about
related talks
33mTurning Fragmented GPU Clusters Into One Elastic Compute Pool
07mProject Lightning Talk: From Static Slices To Elastic GPUs: Dynamic MIG With HAMI
38mStop Allocating GPUs, Start Delivering Intelligence: An Enterprise...
15mKeynote: From Platforms to AI Factories. Has Kubernetes Solved It?
06mSponsored Keynote: AI-Ready Platforms: Scaling Teams Without Scaling Costs
47mGet off my GPU! Better Sharing of AI Accelerators in OpenStack with Blazar Reservations
36mCapacity Scheduling for Elastic Resource Sharing in Kubernetes
24mTaming the Wild West of Research Computing: How Policies Saved Us a Thousand...