Project Lightning Talk: From Idle to Ideal: Cross‑Cluster GPU Sharing with CoHDI - Takao Indoh

This talk introduces CoHDI (Composable Hardware Disaggregated Infrastructure), a sandbox project focused on optimizing GPU sharing across multiple Kubernetes clusters. The speaker, Takao Indoh, discusses the rising demand for AI and HPC workloads amidst the increasing costs and scarcity of GPUs and memory. Traditionally, Kubernetes operates at the cluster boundary, which presents obstacles when one cluster is idle while another requires resources. The presentation explores how CoHDI integrates with Dynamic Resource Allocation and the scheduler to effectively manage GPU resources, enabling real-time attachment and detachment to improve efficiency, reduce costs, and enhance scalability without the need for application rewrites.