State of Checkpoint/Restore in Kubernetes – Now With GPUs
summary
CRIU-based container checkpoint/restore is implemented end-to-end using CRIU + runc + CRI-O/containerd and lands in Kubernetes as an alpha in 2022 and beta in 2024, now running with GPU support via vendor tooling. A live demo migrates an LLM on an AWS A10 GPU: cold start ~38 seconds, restore completes faster. Checkpoints are tar archives converted to OCI images, transfer and GPU state extraction (NVIDIA/AMD drivers) are the bottlenecks being optimized. Kubernetes alpha 2022 beta 2024 GPU checkpointing with NVIDIA/AMD tooling
AI-written summary · about
chapters (7)
- 0:00Introduction and background
- 4:58Checkpoint restore in user space
- 7:24Forensic container checkpointing
- 10:58Kubernetes integration and GPU support
- 16:32Live migration demo
- 21:51Future plans and working group
- 26:29Summary and closing
related talks
39mKubernetes and Checkpoint/Restore
26mKubernetes and Checkpoint Restore
11mLightning Talk: Enabling Hot Restart of Stateful Applications Including GPU-Accelerate...
32mForensic Container Checkpointing
28mContainer Checkpoint/Restore at Scale for Fast Pod Startup Time
39mContainer Live Migration
32mForensic container checkpointing and analysis
30mExplain How Kubernetes Works With GPU Like I’m 5