David Young
David Young
3 min read

Categories

Tags

Recently I helped a Patron to setup an instance of my Emby docker-swarm recipe, but with the extra bonus of having all transcoding done using his GPU. Note that this would work equally well for Plex, or any other “Dockerized” application which would, technically, support GPU processing.

Table of Contents

  1. What’s the big deal about accessing a GPU in a docker container?
  2. How do I enable GPU transcoding with Emby / Plex under docker?
    1. Ubuntu
    2. RedHat / CentOS
    3. Make nvidia-docker the default runtime
  3. Launching a container with docker-nvidia GPU support
  4. Monitoring docker-nvidia GPU usage (with Munin)

What’s the big deal about accessing a GPU in a docker container?

Normally, passing a GPU to a container would be a hard ask of Docker - you’d need to:

  1. Figure out a way to pass through a GPU device to the container,
  2. Have the (quite large) GPU drivers installed within the container image and kept up-to-date, and you’d
  3. Loose access to the GPU from the host platform as soon as you launched the docker container.

Fortunately, if you have an NVIDIA GPU, this is all taken care of with the docker-nvidia package, maintained and supported by NVIDIA themselves.

There’s a detailed introduction on the NVIDIA Developer Blog, but to summarize, nvidia-docker is a wrapper around docker, which (when launched with the appropriate ENV variable!) will pass the necessary devices and driver files from the docker host to the container, meaning that without any further adjustment, container images like emby/emby-server have full access to your host’s GPU(s) for transcoding!

How do I enable GPU transcoding with Emby / Plex under docker?

If you want to learn - read the NVIDIA Developer Blog entry.

If you just want the answer, follow this process:

  1. Install the latest NVIDIA drivers for your system
  2. Have a supported version of Docker
  3. Install nvidia-docker2 (below)

Ubuntu

# Add the package repositories
curl -s -L https://nvidia.github.io/nvidia-docker/gpgkey | \
  sudo apt-key add -
distribution=$(. /etc/os-release;echo $ID$VERSION_ID)
curl -s -L https://nvidia.github.io/nvidia-docker/$distribution/nvidia-docker.list | \
  sudo tee /etc/apt/sources.list.d/nvidia-docker.list
sudo apt-get update

# Install nvidia-docker2 and reload the Docker daemon configuration
sudo apt-get install -y nvidia-docker2
sudo pkill -SIGHUP dockerd

# Test nvidia-smi with the latest official CUDA image
docker run --runtime=nvidia --rm nvidia/cuda nvidia-smi

RedHat / CentOS

# If you have nvidia-docker 1.0 installed: we need to remove it and all existing GPU containers
docker volume ls -q -f driver=nvidia-docker | xargs -r -I{} -n1 docker ps -q -a -f volume={} | xargs -r docker rm -f
sudo yum remove nvidia-docker

# Add the package repositories
distribution=$(. /etc/os-release;echo $ID$VERSION_ID)
curl -s -L https://nvidia.github.io/nvidia-docker/$distribution/nvidia-docker.repo | \
  sudo tee /etc/yum.repos.d/nvidia-docker.repo

# Install nvidia-docker2 and reload the Docker daemon configuration
sudo yum install -y nvidia-docker2
sudo pkill -SIGHUP dockerd

# Test nvidia-smi with the latest official CUDA image
docker run --runtime=nvidia --rm nvidia/cuda nvidia-smi

Make nvidia-docker the default runtime

You could stop here, and manage your containers using the nvidia-docker runtime. However, I like to edit /etc/docker/daemon.json, and force nvidia-docker to be used by default, by adding:

"default-runtime": "nvidia",

And then restarting docker with sudo pkill -SIGHUP dockerd

Launching a container with docker-nvidia GPU support

Even with the default nvidia runtime, the magic GPU support doesn’t happen unless you launch a container with the NVIDIA_VISIBLE_DEVICES=all environment variable set. (Thanks to @flx42 for clarifying this for me)

The advantage to adding the default-runtime argument above, is that you can now deploy your Emby/Plex/Whatever app under swarm exactly as usual, but gain all the benefits of having your GPU available to your app!

Monitoring docker-nvidia GPU usage (with Munin)

Since nvidia-docker now exposes the GPU to the container while still keeping it available to the host OS, you can run nvidia-smi on the host to monitor how the GPU is performing.

I’ve been working on a recipe for a Munin server, so I added the nvidia_ wildcard plugin to monitor nvida GPU stats. Being a “wildcard”, it’s necessary to copy the plugin to /usr/share/munin/plugins, and then symlink it mulitple times to /etc/munin/plugins/ as below:

ln -s /usr/share/munin/plugins/nvidia_gpu_ /etc/munin/plugins/nvidia_gpu_temp
ln -s /usr/share/munin/plugins/nvidia_gpu_ /etc/munin/plugins/nvidia_gpu_mem
ln -s /usr/share/munin/plugins/nvidia_gpu_ /etc/munin/plugins/nvidia_gpu_fan
ln -s /usr/share/munin/plugins/nvidia_gpu_ /etc/munin/plugins/nvidia_gpu_power
ln -s /usr/share/munin/plugins/nvidia_gpu_ /etc/munin/plugins/nvidia_gpu_utilization