CoHDI: A CNCF Sandbox Project for Dynamic Hardware Composability in Kubernetes
summary
CoHDI is a CNCF Sandbox project that implements dynamic attachment and detachment of GPUs and other PCI/CXL devices into Kubernetes using three components: Composable DRA Driver, Dynamic Device Scaler, and Composable Resource Operator. A pod pending on a ResourceClaim triggers the scaler to call a fabric manager to attach devices (example: cluster scales from 2 to 4 GPUs), the driver publishes resource slices so the pod then schedules. Current demos run with NVIDIA GPUs, NICs and CXL memory, two vendor implementations (FSAS, NEC); topology‑aware allocation and low‑latency (APN/optical) networks remain open issues.
AI-written summary · about
chapters (7)
- 0:00CoHDI CNCF Sandbox intro
- 3:16Fixed server limits for AI
- 7:21CoHDI dynamic hardware
- 9:36CoHDI stack and workflow
- 16:50Global deployment with CoHDI
- 19:47AI infra vision challenges
- 23:06Q&A on GPU services
related talks
41mDynamic GPU Provisioning in Kubernetes With Composable Disaggregated...
04mFrom Idle to Ideal: Cross‑Cluster GPU Sharing with CoHDI
38mDynamic Scaling of GPUs for Container Apps with Composable Disaggregated...
34mDeclarative Device Virtualization: Orchestrating GPUs & Hardware in Cloud Native Environments
05mTowards a Sustainable World Realized by Kubernetes and Composable...
30mRunning AI Workloads in Containers and Kubernetes
05mHami: Dynamic, Smart, Stable GPU-Sharing Middleware In Kubern...
53mDeclarative Device Virtualization: Orchestrating GPUs & Hardware in Cloud Native Environments