nvidia-docker-compose

nvidia-docker-compose is a simple python script that wraps docker-compose to allow docker-compose to work with GPU enabled Docker containers as made available with nvidia-docker!

Dependencies

nvidia-docker-compose requires following dependencies to be installed on the system:

• Docker engine
• nvidia-docker

It also depends on the docker-compose, PyYAML and Jinja2 Python packages, which would be installed automatically during the installation step described below.

nvidia-docker-compose and nvidia-docker2

As some of you may know, NVIDIA is working on the release of nvidia-docker2 that integrates much more tightly with Docker infrastructure (more details can be found at the master branch of their project repo https://github.com/NVIDIA/nvidia-docker). One of the huge benefit of this new integration is that you would no longer need nvidia-docker-compose to launch docker-compose with GPU capacity. Refer to the issue #23 for how you could configure nvidia-docker2 to work with docker-compose.

Before you install

nvidia-docker-compose depends on nvidia-docker to properly function and above all, it depends on all extra Docker volumes that are automatically created when you run nvidia-docker. Before you install and run nvidia-docker-compose, please make sure to test run nvidia-docker at least once to ensure that all volumes are set up and are functioning correctly. In particular, I recommend that you run the following command:

$ nvidia-docker run --rm nvidia/cuda nvidia-smi

If this runs and properly lists all available GPUs on your machine, then you are ready to proceed! If not, please refer to nvidia-docker documentation and helps to make sure that it functions properly before using nvidia-docker-compose.

Installing

To install the script, simply run:

$ pip install nvidia-docker-compose

If you are using system Python, it may be necessary to run the above command with sudo upfront.

Using nvidia-docker-compose

The nvidia-docker-compose is a drop-in replacement for the docker-compose. Simply run as you would run docker-compose:

$ nvidia-docker-compose ...

Depending on how your system is configured, you may need to run the script with sudo (i.e. if you usually need sudo to run docker, you will need sudo).

Running nvidia-docker-compose generates a new YAML config file nvidia-docker-compose.yml locally. It is safe to delete this file in-between usages and I recommend you add this to your .gitignore file if you are going to use nvidia-docker-compose within a Git repository. Once generated, you can also use the nvidia-docker-compose.yml directly to launch GPU enabled containers directly with the standard docker-compose. You can do so as:

$ docker-compose -f nvidia-docker-compose.yml ...

Running flexibly on multi-GPU setup

When working on multi-GPU setup, you would often want to run separate container for each GPU or at least limit the visibility of GPUs to only specific Docker containers. If you are not afraid to dig in, you would discover that you can control visibility of GPUs to each container by selectively including /dev/nvidia* under the devices section (i.e. /dev/nvidia0 for the first GPU, and so on) . However, doing this manually would mean that you will have to interfere with the function of nvidia-docker and nvidia-docker-compose, and previously there was no natural way to specify which service in the docker-compose.yml should be run with which GPUs. This is further complicated by the fact that different machine would have different numbers of GPUs, and thus keeping a service with /dev/nvidia4 under devices section on a 2 GPU machine could cause an error.

Specifying GPU target

New from version 0.4.0 nvidia-docker-compose now allows you to specify which GPU a specific service should be run with by including /dev/nvidia* under the devices heading. As in the following

version: "2"
services
  process1:
    image: nvidia/cuda
    devices:
      - /dev/nvidia0
  process2:
    image: nvidia/cuda
    devices:
      - /dev/nvidia1
      - /dev/nvidia2

The service process1 will now only see the first GPU (/dev/nvidia0) while the service process2 will see second and third GPU (/dev/nvidia0 and /dev/nvidia1). If you don't specify any /dev/nvidia* under devices section, the service will automatically see all available GPUs as have been the case previously.

Although this feature will allow you to finely control which service sees which GPU(s), it is still rather inflexible as will require you to adjust the docker-compose.yml per computer device. This is precisely where the Jinja2 templating can help you!

Using Jinja2 in docker-compose.yml file

New from version 0.4.0

To support the relatively common use case of wanting to launch as many compute containers (with the same configuration) as the number of GPUs available on the target machine, nvidia-docker-compose now supports use of Jinja2. Combined with the ability to specify GPU targeting, you can now write docker-compose config that adapts flexibility to the GPU availability. For an example if you prepare the following template and save it as docker-compose.yml.jinja:

version: "2"
services:
  {% for i in range(N_GPU) %}
  notebook{{i}}:
    image: eywalker/tensorflow:cuda
    ports:
      - "300{{i}}:8888"
    devices:
      - /dev/nvidia{{i}}
    volumes:
      - ./notebooks:/notebooks
  {% endfor %}

and specify the target Jinja2 template with -t/--template flag when you run:

$ nvidia-docker-compose --template docker-compose.yml.jinja ...

It will pick up the Jinja template, process it and expand it to the following docker-compose.yml:

version: "2"
services:
  notebook0:
    image: eywalker/tensorflow:cuda
    ports:
      - "3000:8888"
    devices:
      - /dev/nvidia0
    volumes:
      - ./notebooks:/notebooks
  notebook1:
    image: eywalker/tensorflow:cuda
    ports:
      - "3001:8888"
    devices:
      - /dev/nvidia1
    volumes:
      - ./notebooks:/notebooks
  notebook2:
    image: eywalker/tensorflow:cuda
    ports:
      - "3002:8888"
    devices:
      - /dev/nvidia2
    volumes:
      - ./notebooks:/notebooks

on a 3 GPU machine. The Jinja variable N_GPU automatically reflects the available number of the GPUs on the system. This docker-compose.yml is then processed by nvidia-docker-compose just like any other config file to launch GPU enabled containers.

Generating Compose File Only

If you want to generate GPU-enabled compose file for later use, -G/--generate flag will make nvidia-docker-compose exit after generating the compose file without running docker-compose.

$ nvidia-docker-compose -G ...

Additional command line options

For additional configurations such as specifying alternate nvidia-docker-plugin host address, alternate target docker compose file name (instead of the default nvidia-docker-compose.yml), refer to the command line help at:

$ nvidia-docker-compose -h

How it works

nvidia-docker-compose is a simple Python script that performs two actions:

• parse docker-compose config file (defaults to docker-compose.yml) and creates a new config YAML nvidia-docker-compose.yml with configurations necessary to run GPU enabled containers. Configuration parameters are read from nvidia-docker-plugins.
• run docker-compose with the newly generated config file nvidia-docker-compose.yml