Running a remote Jupter Notebook Server on Wynton HPC

Instructions on running a jupyter notebook instance on UCSF Wynton High Performance Compute Cluster. Since wynton access utilizes SSH access, familiarize yourself on using SSH. On Mac/Linux, this can easily be access via terminal. This is enabled by default on newer versions of Windows 10. Some older versions require extra steps.

Wynton Access

You will need to first obtain access to the UCSF Wynton server. This can be done here.

After obtaining access, use the procedure to log in. This is done through SSH. In Mac/Linux, SSH is accessed via Terminal. In Windows, you're going to want to use Powershell.

I will be using log2.wynton.ucsf.edu as the main access point for this notebook. You'll generally have access to more powerful hardware on the development nodes. The way these are accessed is by entering ssh NODE_NAME only after successfully accessing the login node.

e.g. the commands could be entered in the order of:

ssh [email protected]
ssh NODE_NAME

Node names and specs can be found here.

Nodes are selected based on hardware needs. General coding will benefit from the Xeon E5430 CPU found on dev1, but using large datasets will probably require dev2 or dev3 which have larger /scratch designation. Deep learning (and other GPU-dependent tasks) should be done on gpudev1.

SSH Configuration

Storing a public-private ssh key pair will save a lot of headaches in the future. What this allows us to do is to login to the server without having to enter a password. This is especially useful because accessing the development nodes requires you to enter your password twice. Instructions are pulled from here.

1. From Local Machine: ssh-keygen -t rsa
2. Enter file in which to save the key

• Press ENTER (Leave Default)

3. Enter passphrase

• Press ENTER (Leave Default)

4. Enter passphrase again

• Press ENTER (Leave Default)

5. MacOS:

   curl https://raw.github.com/beautifulcode/ssh-copy-id-for-OSX/master/ssh-copy-id.sh -o \n
   /usr/local/bin/ssh-copy-id
   
   ssh-copy-id -i ~/.ssh/id_rsa.pub [email protected]

   Windows:

   cat ~/.ssh/id_rsa.pub | ssh [email protected] "cat >> ~/.ssh/authorized_keys"

• If applicable: Are you sure you want to continue connecting? (yes/no) Type yes
• Enter password when prompted

This allows us to access the login node without being prompted for a password.

Access a node by entering ssh NODE_NAME after logging into the node. If you log in, skip to the Setting up Jupyter Notebook Server section.

If you are prompted for a password, you will need to repeat the process, this time pairing the login node to the dev:

1. ssh [email protected]
2. ssh-keygen -t rsa
3. Enter file in which to save the key

• Press ENTER (Leave Default)

4. Enter passphrase

• Press ENTER (Leave Default)

5. Enter passphrase again

• Press ENTER (Leave Default)

6. ssh-copy-id -i ~/.ssh/id_rsa.pub USERNAME@NODE_NAME

• Enter password when prompted

Make sure the terminal window says [USERNAME@wynlog2 ~] during this set of steps. It is possible that you get logged out. Remember, we are creating that password pairing between your account's login node and the development node.

Setting up Jupyter Notebook Server

Package Installation

Now that we have the password pairings established, we can finally begin to install our python packages. The python installation is barebones so we need to install EVERY package we desire manually.

Login with

ssh -t [email protected] 'ssh USERNAME@NODE_NAME'

Update pip

In the development node, enter:

pip install --user --upgrade pip
pip3 install --user --upgrade pip

Install your favorite packages

UCSF has Python 2 and 3 installed, and very little else. You will need to install any package you intend on using via pip3 install --user PACKAGE NAME" You can copy and paste the snippit below for the essentials. I will be doing these installations on Python 3, but if you would like to use Python 2.x, simply use pip instead of pip3.

In the development node, enter:

pip3 install --user numpy scipy matplotlib pandas scikit-learn

Install Jupyter Notebooks

Now we're at the meat and potatoes of this long endeavor. You will be installing jupyter notebooks and configuring it to run as a server which you can access remotely.

In the development node, enter:

pip3 install --user jupyter

I have included a config file which you can use. Download or clone this repo and enter the following command to upload jupyter_notebook_config.

From local machine (Type exit in Terminal):

scp CONFIG_FILE_LOCATION [email protected]:"~/.jupyter"

Fulll instructions to set up a jupyter notebook server can be found here. You may configure your own config file if there are other options you'd like to set.

Accessing the notebook server

Here's the moment you've all been waiting for. We will now initiate the notebook instance and access the server from our local machines. This is done by connection ports between your machine and the login node, and then the login node with the development node.

First decide on a port number. In the AppleScript I've attached, it generates a 4 digit number randomly. Try to avoid common ones like 8888 and 8880. Open a fresh terminal window and type:

ssh -N -f -L localhost:PORT:localhost:PORT [email protected];
ssh -t [email protected] 'ssh -N -f -L localhost:PORT:localhost:PORT USERNAME@NODE_NAME';
ssh -t [email protected] 'ssh USERNAME@NODE_NAME'
jupyter notebook --port=PORT

Now just open your browser and type localhost:PORT in the address bar to access the notebook server.

You will need to do this every time you intend on accessing the notebook. It is fairly simple to write a script to automate this. I've attached a launchable applescript for Mac in the git which you'll need to modify with your own credentials.

The first time you access it, you will be asked to generate a password to access the notebook.

The key can be found in the terminal window, as seen below:

From here, you can upload files via the GUI and work on your projects.