GPU-jupyterhub

This jupyterhub implementation allows for Nvidia GPU access using the nvidia-docker-2 container runtime.

Requirements

• A cuda driver must be installed on the host system, you can check this by running nvidia-smi in the terminal.
• Docker 19.03 or higher.
• Docker compose 1.25.5 or higher.

I've personally found the DigitalOcean Tutorial to be the most reliable. Make sure to change the version number to 1.25.5!

• The nvidia-container-runtime needs to be installed:

sudo apt-get install nvidia-container-runtime

• Nvidia docker2 needs to be installed see their Github for instructions.

Installation

Preparation

To make runtime: nvidia work we need to change our /etc/docker/daemon.json to the following:

{
    "runtimes": {
        "nvidia": {
            "path": "/usr/bin/nvidia-container-runtime",
            "runtimeArgs": []
        }
    }
}

Building our notebook containers

We can now build our notebook containers with:

#cd notebooks/{notebook-folder}
#docker build -t {notebook-folder-name} .
Example
cd notebooks/base-notebook
docker build -t "base-notebook" .
cd .. cd notebooks/minimal-notebook docker build -t "minimal-notebook"
And so on

Building the hub

Note: Make sure to change the userlist file to include your Github username.

# Make sure to do this in the root of the repo*
docker-compose up --build

Common Issues

• Volume jupyterhub-db-data or jupyterhub-data not found.

docker volume create --name="jupyterhub-data"

• Network jupyterhub-network not found.

docker network create "jupyterhub-network"

• No such file or directory: '/data/jupyerhubcookiesecret'

Run the following command whilst replacing $DATAVOLUMECONTAINER with the actual path.

openssl rand -hex 32 > {$DATAVOLUMECONTAINER}/jupyterhubcookiesecret