Utku Evci

Utku Evci

Feed your head

SSH Port Forwarding and Jupyter Notebook

3 minute read

In the last few years Jupyter notebooks became a vital part of my coding routine. I first started using jupyter notebooks as a python interpreter and run it locally to try out some basic things. Than I learned that one can run the notebook in a server and open it locally in the browser. With this one can use the power of a remote server as if working on a jupyter notebook locally.

In this post I will explain three things

  • How to start a jupyter notebook on a server using local forwarding.
  • How to start jupyter notebook-job.

For other things like exchanging key, ssh aliasing&tunneling check this.

Jupyter Notebooks with Basic Local Forwarding.

When you run

jupyter notebook

a notebook is created on the default port, which is http://localhost:8888/ on my Mac. One can use --port **** flag to specify the port of the local host. If we want to run the notebook on a server through ssh, we can use Local Forwarding which is the -L flag. Lets assume we have a server named hostOne defined in ~/.ssh/config/. For more information about how to do that you can check out my ssh notes).

ssh -L 8000:localhost:8888 hostOne
jupyter notebook --port 8888

Running the two lines above would start a notebook on the server hostOne at port 8888. And local forwarding connects the localhost at port 8000 to the port 8888 of the server.

Running a Jupyter notebook on a Slurm Cluster

Here we need understand ssh forwarding&tunneling. Even though it seems pretty complicated at the beginning, this answer explains it pretty well. Great vis!. When a job is started, it runs on a different address and we need to create a tunnel from the job-server to the cluster we are connecting(the one we submit job in). In our example the cluster is NYU’s prince and it uses Slurm scheduler. What we need to do is first submit a job, which does the forwarding, starts the jupyter notebook and informs us. The job would be like following (inspired from here). Lets name it jupyterGPU.job.

#!/bin/bash
#SBATCH --job-name=jupyterTest
#SBATCH --nodes=1
#SBATCH --gres=gpu:1
#SBATCH --mem=8GB
#SBATCH --time=2:00:00

#Load necessary modules
module purge
module load python/intel/2.7.12

#Go to the folder you wanna run jupyter in
cd $HOME

#Pick a random or predefined port
#port=$(shuf -i 6000-9999 -n 1)
port=8765
#Forward the picked port to the prince on the same port. Here log-x is set to be the prince login node.
/usr/bin/ssh -N -f -R $port:localhost:$port log-0
/usr/bin/ssh -N -f -R $port:localhost:$port log-1

#Start the notebook
jupyter notebook --no-browser --port $port

Than we submit the job

sbatch run-jupyter.sbatch
#Submitted batch job 953167

When it is completed we have the log named slurm-953167.out with output similar to this.

[I 17:41:05.035 NotebookApp] Writing notebook server cookie secret to /state/partition1/job-953167/jupyter/notebook_cookie_secret
[I 17:41:06.051 NotebookApp] Serving notebooks from local directory: /home/ue225
[I 17:41:06.052 NotebookApp] 0 active kernels 
[I 17:41:06.052 NotebookApp] The Jupyter Notebook is running at: http://localhost:8765/?token=33703785bdb10cadf4ab0645002ab373e8e966b114f05c11
[I 17:41:06.052 NotebookApp] Use Control-C to stop this server and shut down all kernels (twice to skip confirmation).

All we need to at this point is to run

ssh -L 8765:localhost:8765 ue225@prince

and open http://localhost:8765/?token=33703785bdb10cadf4ab0645002ab373e8e966b114f05c11 in your browser.

You May Also Enjoy