Ctrl+K

8.3. Working with Servers

8.3. Working with Servers#

In a real deep learning project, we often need to run our code on a server computer with powerful GPUs.

Connection#

We can use the Secure Shell Protocol (SSH) to connect to a server computer:

ssh <USER_NAME>@<SERVER_IP>

After entering the password, you have made a connection to the server computer. You have access to the terminal of the server computer where you can execute different commands.

Jupyter#

You can start a jupyter notebook/lab on the server and open it on your local browser.

On the server run:

jupyter notebook --no-browser --port=<REMOTE_PORT>

This is similar to how we start a jupyter notebook/lab on a local machine with an extra argument of --no-browser that starts the notebook without opening a browser.

Note that the port <REMOTE_PORT> you selected might not be the one that gets assigned to you (e.g., in case it’s already being used).

Once jupyter notebook has started, it will show you an URL with a security token that you will need to open the notebook in your local browser.

On your local terminal:

ssh -L <LOCAL_PORT>:localhost:<REMOTE_PORT> <REMOTE_USER>@<SERVER_IP>

This command links the <REMOTE_PORT> to the specified <LOCAL_PORT>.

Once the connection is set up, you can open jupyter notebook in your browser by entering:

http://localhost:<LOCAL_PORT>/

You may be asked to enter a token (see above).

TensorBoard#

Similar to Jupyter, you can start the tensorboard on the server and open it on your local browser. There are two options.

“No particular address” placeholder#

On the server run:

tensorboard --host 0.0.0.0 --logdir <LOG_DIR> --port <REMOTE_PORT>

On the browser of your local computer:

http://<SERVER_IP>:<REMOTE_PORT>/

Some servers’ firewall block this method. If this is the case, use the local port forwarding.

Local port forwarding#

On the server run:

tensorboard --logdir <LOG_DIR> --port <REMOTE_PORT>

On local terminal:

ssh -L <LOCAL_PORT>:localhost:<REMOTE_PORT> <REMOTE_USER>@<SERVER_IP>

On the browser of your local computer:

http://localhost:<LOCAL_PORT>/

Tmux#

In most use-case scenarios, you want:

  • to run a script for several hours,

  • to run several scripts at the same time.

tmux facilitates this functionality.

  • To create a new session run tmux new -s <SESSION_NAME>.

  • To detach from a session press ctr+b and then d.

  • To attach to a session press tmux a -t <SESSION_NAME>.

  • To easily move across sessions press ctr+b and then s.

Online resources:

Contents