From Idle to Ideal: Cross‑Cluster GPU Sharing with CoHDI

on YouTube ↗04m

" summary

CoHDI presents a composable hardware system that attaches/detaches GPUs on demand from a centrally managed shared GPU and memory pool, integrating with Dynamic Resource Allocation and the Kubernetes scheduler to reassign devices across clusters. Prototype code v0.0.1.1 and a composable infrastructure emulator exist, currently tracking scheduler changes for Kubernetes 1.36 as a beta. Claimed benefits are reduced GPU idle time and lower cost by moving devices between inference and training workloads at runtime.

AI-written summary · about

" related talks

• 30mCoHDI: A CNCF Sandbox Project for Dynamic Hardware Composability in KubernetesNaoki Oguchi · KubeCon + CloudNativeCon Japan 2026
• 41mDynamic GPU Provisioning in Kubernetes With Composable Disaggregated...Jin Hase, Tsubasa Watanabe · Open Source Summit Japan + AI_dev: Open Source GenAI & ML Summit + Automotive Linux Summit 2025
• 05mProject Lightning Talk: CoHDI: Dynamic Hardware Composition for AI‑Era Kubernetes…N. Oguchi · KubeCon + CloudNativeCon Japan 2026
• 34mDeclarative Device Virtualization: Orchestrating GPUs & Hardware in Cloud Native EnvironmentsSamrat Priyadarshi, Anmol Krishan Sachdeva · AI_dev Europe 2025
• 31mAI, CERN, and the Quest for GPU Custody: How CERN Leverages DRA for Efficient GPU SharingDiana Gaponcic, Jan-Philip Gehrcke · Cloud Native + Kubernetes AI Day Europe 2025
• 10mLightning Talk: K8s Issue #52757: Sharing GPUs Among Multiple ContainersXiao Zhang · Cloud Native AI + Kubeflow Day 2026
• 38mDynamic Scaling of GPUs for Container Apps with Composable Disaggregated...Lei Zhang, Jin Hase · Open Source Summit Japan + AI_dev 2024
• 38mImproving GPU Utilization using KubernetesMaulin Patel, Pradeep Venkatachalam · KubeCon + CloudNativeCon Europe 2022