Server Setup Log

Initial Setup

The server was initially set up by installing Ubuntu 22.04.

Requirements

The primary requirements for the server were to support:

• OpenMPI
• OpenMP
• CUDA

Solution

To meet these requirements, a JupyterHub instance was set up.

• We utilized DockerSpawner to manage user environments.
• The setup includes full support for GPU acceleration, ensuring users can leverage CUDA for deep learning tasks.

Build Process

1. Download Driver: The nvhpc_2025_253_Linux_x86_64_cuda_12.8.tar.gz driver was downloaded first.
2. Build Image: The Docker image was then built, incorporating the downloaded driver.

DockerSpawner Implementation

Why DockerSpawner? We chose DockerSpawner to manage user instances because it provides:

• Isolation: Each user runs in their own isolated Docker container, preventing conflicts between users' processes and dependencies.
• Consistency: Every user starts with the exact same environment (defined by our Docker image), ensuring reproducibility for deep learning tasks.
• Flexibility: It allows us to easily update the environment by rebuilding the image without affecting the host system.

How it Works

1. When a user logs in to JupyterHub, the Hub calls DockerSpawner.
2. DockerSpawner communicates with the Docker daemon to create a new container for that specific user.
3. This container runs the jupyterhub-singleuser server.
4. JupyterHub then proxies traffic from the user's browser to their specific container.

References

• JupyterHub Documentation
• DockerSpawner Documentation