Run Jupyter Server with GPU Access on HPCs

You can run a Jupyter notebook/lab server that has GPU access inside your container and connect to it from your local computer. The following steps describe the details:

1. Set up ssh connection with DGX or DLS as described in Connect to HPCs

After completing the setup process, you should be able to connect to DGX or DLS with ssh.

2. Add port forwarding

In your ssh config file, inside DGX and DLS environments, add the following line:

LocalForward 56789 127.0.0.1:56789

To avoid conflict with other people and other applications, the port number 56789 should be replaced by a port number in [49152, 65535]. The modified config file should be similar to this:

Host jump-box
      HostName eniac.cs.hunter.cuny.edu
      User eniac_username
      IdentityFile ~/.ssh/id_rsa (path/to/your/id_rsa)
      ServerAliveInterval 240

   Host dgx
      HostName 146.95.214.135
      User dgx_username
      IdentityFile ~/.ssh/id_rsa (path/to/your/id_rsa)
      ProxyCommand ssh -q -W %h:%p jump-box
      LocalForward 56789 127.0.0.1:56789
      ServerAliveInterval 240

   Host dls
      HostName 146.95.128.169
      User dls_username
      IdentityFile ~/.ssh/id_rsa (path/to/your/id_rsa)
      ProxyCommand ssh -q -W %h:%p jump-box
      LocalForward 56789 127.0.0.1:56789
      ServerAliveInterval 240

3. Start container with GPU port published to host

Add an additional -p argument when starting your container:

nvidia-docker run --name containerName --network=host -p 56789:56789 -it -v localDir:containerDir imageRepository:tag

4. Insider your container, start Jupyter server

Start your Jupyter notebook/lab server with:

jupyter notebook --ip 0.0.0.0 --port 56789 --allow-root

or

jupyter lab --ip 0.0.0.0 --port 56789 --allow-root

To run the server in the background, so you can still do other things within your container, start the server with nohup:

nohup jupyter notebook --ip 0.0.0.0 --port 56789 --allow-root > jupyter.log &

or

nohup jupyter lab --ip 0.0.0.0 --port 56789 --allow-root > jupyter.log &

Notice the & symbol at the end of the commands.

5. Connect to the server

On your local machine, type localhost:56789 in your browser address bar and hit enter. If this is your first connection, you will see a page similar to:

You can find the token in your jupyter.log file. The token should look like this:

Copy and paste the token into the text bar and click log in. Now you should be able to enjoy the convenience of Jupyter notebook/lab on our DGX/DLS with the power of GPUs.

6. (Optional) Set up a password for your Jupyter notebook

If you do not like the long token string, you can set up a password for your notebook. To do that, before starting the server, you need to create a config file for Jupyter:

jupyter notebook --generate-config

then create a password:

jupyter notebook password

Type in your password twice, then an encrypted password will be saved into the config file you created above. Now you can start the Jupyter server as in step 4. When you connect to the server started this way, the login page will only ask you for the password.