Design Advantages and Implementation Mechanism of nvidia-docker 2.0

In the first part, the article reviews the limitations of the original nvidia-docker 1.0, which tightly couples with the Docker runtime, lacks flexibility, cannot leverage other runtimes or Docker Compose, and does not treat GPUs as schedulable resources.

It then introduces nvidia-docker 2.0, which uses the NVIDIA Container Runtime, nvidia-container-runtime, libnvidia-container, and runc, modifying Docker's /etc/docker/daemon.json to switch the default runtime.

Key components are described: nvidia-docker2.0 (a simple package that changes Docker configuration), nvidia-container-runtime (adds a pre‑start hook to invoke libnvidia-container), libnvidia-container (provides a library and CLI to expose NVIDIA GPUs to containers), and runc (the OCI runtime used by containerd).

The article explains the container creation flow for a normal container ( docker → dockerd → docker‑containerd‑shim → runc → container‑process ) and for a GPU‑enabled container ( docker → dockerd → docker‑containerd‑shim → nvidia‑container‑runtime → container‑process ), highlighting that the NVIDIA runtime replaces the default runtime and uses a hook to check the NVIDIA_VISIBLE_DEVICES environment variable before invoking libnvidia‑container.

Additional background on containerd is provided, outlining its responsibilities such as lifecycle management, image pull/push, storage management, network handling, and invoking runc.

The article concludes with references to relevant GitHub repositories and documentation for further exploration.