PGAP 2023 Software

Software installation for GPU course

• Day one
• Day two
• 2023 live lectures

Software installation.

To actively participate in this course, you need the following software:

• NVIDIA CUDA drivers
• NVIDIA CUDA toolkit
• Anaconda / miniconda

Installation instructions for Ubuntu:

1. Install CUDA drivers and toolkit. The following should be sufficient:

sudo apt-get update
sudo apt-get -y install nvidia-cuda-toolkit
restart

2. Then continue to install Anaconda by following the instructions here: https://docs.anaconda.com/anaconda/install/linux/

After these two are installed, you should be able to create a new conda environment. Start by downloading the following YML evironment file gpu_env.yml. In a terminal, execute the following:

conda env create --name gpu_env --file gpu_env.yml
conda activate gpu_env

You should now be able to run jupyter notebooks with pycuda. Try it out with

conda activate gpu_env
jupyter-lab

and a browser window with jupyter lab should appear.

Accessing jupyter on a remote server

Let us assume that we have an SSH server that has a GPU, and we want to access jupyterlab on that machine. The following is a recipe for that. Let us name the two machines laptop and gpu-server. The procedure can be listed as follows:

1. SSH to gpu-server as usual. Open a terminal, and ssh to the server:

ssh username@gpu-server

2. Start jupyterlab (without browser). In the same terminal, start jupyterlab:

jupyter lab --no-browser

Make a note of the port that jupyterlab starts on and the token. You will need it for the next point. They have the following format:

http://localhost:<PORT>/lab?token=<TOKEN-CODE>

3. Open ssh tunnel which connects a port on laptop with the jupyterlab port on gpu-server. In a new terminal, execute the following:

ssh -L 8000:localhost:8888 gpu-server

In this command, you need to replace 8888 with the port number from jupyterlab in the previous point.

4. Open browser to interact with jupyterlab on the gpu-server. Open http://localhost:8000/lab?token=<TOKEN-CODE> in your browser, and you should be able to access jupyterlab. Make sure you replace <TOKEN-CODE> with what jupter lab wrote to the terminal in the step above.