SDSC Expanse Notebook: CUDA_GPU_NVIDIA

This guide provides instructions for Expanse users to run CUDA notebooks on GPU nodes. Code authored by Abe Stern, NVIDIA.
Listof Content

• Launch Galyleo
• Import Module
• Environment Modules
• Install Modules
• Location
• Short Description
• Sources
• Submit Ticket

Import Module:

• numba
• math
• numpy
• cuda
• vectorize
• cuda

Launch Galyleo

For specific information about launching Galyleo, please refer to this GitHub repository.

Environment Modules

By utilizing –env-modules, we can load any software installed on Expanse. Since CUDA cannot run directly on a CPU, we need to load GPU modules to run in a GPU environment on Expanse.

Instead of using modules, we can also use conda environments supported by Galyleo, utilizing the --conda-env option. The following command line launches a conda environment on Expanse for parallel GPU processing. For more information, please refer to this GitHub repository.

galyleo launch --account sds173 --partition gpu-shared --cpus 10 --memory 92 --gpus 1 --time-limit 00:30:00 --conda-env df-parallel-gpu --conda-yml "/home/(username)/df-parallel/environment-gpu.yml" --mamba

Install Modules

To run cuda_gpu_nvidia notebooks, no additional package installation is required.

Location

CUDA_GPU_NVIDIA
├── cuda_gpu_nvidia_computing_pi_solution.ipynb
├── cuda_gpu_nvidia_distance_matrix_solution.ipynb
├── cuda_gpu_nvidia_law_of_cosines_solution.ipynb
├── README.md

Short Description

In the Computing Pi exercise, we will design a CUDA kernel to compute the value of Pi via Monte Carlo. The concepts of writing and invoking CUDA kernels in Numba are introduced.

In the Distance Matrix exercise, we will compute a distance matrix for a synthetic dataset of 3-D molecular geometries. We will learn how to leverage higher-dimensional CUDA thread-block hierarchies.

In Law of Cosines exercise, we will explore GPU Ufuncs which are simple to write, invoke, and are compatible with Numpy Ufuncs. We will learn how to write a simple GPU Ufunc to compute the law of cosines.

Sources

Below are listed a few related readings and presentations. Numba supports CUDA GPU programming by directly compiling a subset of Python code into CUDA kernels and device functions following the CUDA execution model.

In the provided notebooks, a problem is introduced and mostly implemented in Numba. As an exercise, complete the missing lines of code to successfully compute the result.

These notebooks were part of th SDSC HPU User Training Spring 2020 Session Week 4 (01/31/2020) and were presented by Abe Stern, NVIDIA with the topic of GPU accelerated computing with CUDA Python.

Submit Ticket

If you find anything that needs to be changed, edited, or if you would like to provide feedback or contribute to the notebook, please submit a ticket by contacting us at:

Email: consult@sdsc.edu

We appreciate your input and will review your suggestions promptly!