Investigating Checkpoint and Restore for GPU-Accelerated Containers
recording made private on YouTubehttps://youtu.be/khtcxcI4rrg
summary
Demonstrates that CRIU can checkpoint and restore GPU-accelerated containers today using an CRIU plugin for AMD GPUs, checkpoint archives that are essentially memory pages plus changed files, and a live demo migrating a container from Europe to Seattle where GPU load drops from ~77% to 0% after checkpoint and resumes on restore. Uses Podman for the demo and Kubernetes support via the CRI is available but still early. Nvidia support remains unclear and restore performance depends on checkpoint I/O and device vendor tooling.
AI-written summary · about
