Sharing is Caring – Fractional GPU Allocations With MetaGPU Device Plugin
Talkat Kubernetes Batch + HPC Day Europe 2023 · on Tuesday, 2023-04-18· at 13:45· Hall 7, Room A | Ground Floor | Europe Complex·
summary
Build that splits any GPU into 100 meta-GPUs per card, registers capacity to kubelet via a gRPC device plugin and exposes usage with mgctl and Prometheus exporter. It enforces per-share GPU memory by killing over‑limit processes, supports dynamic resharing with zero downtime, and targets older cards that lack MIG. The open‑source project called MetaGPU device plugin runs as a DaemonSet, is production‑tested, and reports allocatable vs allocated so Kubernetes scheduling becomes accurate.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 1:12Linux GPU basics
- 2:16GPU sharing with Docker
- 4:01GPU sharing in Kubernetes
- 6:53GPU sharing without plugins
- 9:24Multi-node GPU challenges
- 11:55Kubernetes device plugins
- 12:56MetaGPU architecture
- 14:02Fractional GPU allocation
- 16:14NVIDIA monitoring limits
- 17:38mgctl for Kubernetes
- 19:26Enforcing GPU memory limits
- 23:04Open source overview
- 23:25MetaGPU features
- 23:47Q&A
related talks
38mImproving GPU Utilization using Kubernetes
30mGPU Sharing at CERN: Cutting the Cake Without Losing a Slice!
10mLightning Talk: K8s Issue #52757: Sharing GPUs Among Multiple Containers
05mLightning Talk: Sharing a GPU Among Multiple Containers
32mScaling AI Workloads with Kubernetes: Sharing GPU Resources Across Multiple Containers
04mLightning Talk: Dynamic GPU Fraction and Sharing With Cloud Native Principle- 32mScaling AI Workloads with Kubernetes: Sharing GPU Resources Across Multiple Containers
05mHami: Dynamic, Smart, Stable GPU-Sharing Middleware In Kubern...