Difference between revisions of "GPU Checkpointing"
m |
m |
||
| Line 7: | Line 7: | ||
=== CUDA Plugin === | === CUDA Plugin === | ||
| − | The checkpointing functionality for CUDA applications is enabled through | + | The checkpointing functionality for CUDA applications is enabled through the [http://github.com/NVIDIA/cuda-checkpoint cuda-checkpoint] utility. This utility allows to transparently checkpoint and restore the CUDA state of a running Linux process. The CUDA plugin integrates this external utility with CRIU to safely pause, checkpoint, and restore processes and containers that use NVIDIA GPUs. It detects whether <code>cuda-checkpoint</code> is available and whether the system has an NVIDIA GPU; if not, the plugin disables itself. |
=== AMDGPU Plugin === | === AMDGPU Plugin === | ||
Revision as of 22:11, 10 November 2025
This page explains how CRIU handles GPU-accelerated workloads, what vendor components are needed (NVIDIA & AMD), and how to use them for processes and containers. For more information, see the CRIUgpu paper.
How CRIU integrates with GPU checkpointing mechanisms?
CRIU checkpoints Linux-kernel resources (e.g., memory, threads, files, sockets). GPU state such as device memory, contexts, and queues that lives outside normal process address space needs special handling, so CRIU relies on vendor-specific plugins.
CUDA Plugin
The checkpointing functionality for CUDA applications is enabled through the cuda-checkpoint utility. This utility allows to transparently checkpoint and restore the CUDA state of a running Linux process. The CUDA plugin integrates this external utility with CRIU to safely pause, checkpoint, and restore processes and containers that use NVIDIA GPUs. It detects whether cuda-checkpoint is available and whether the system has an NVIDIA GPU; if not, the plugin disables itself.