Lehman Garrison
Research Software Engineer
Scientific Computing Core
Flatiron Institute
This repo contains lectures and problem sets from LSST DSFP Session 21 classes IX and X on "High-Performance Python and Parallelization" and "GPU-Accelerated Python". The materials cover a bit of NumPy, Numba, CPU parallelism with Numba, CuPy, and CUDA with Numba. The notebooks should take roughly 1–1.5 hours each to go through. They assume a little bit of Python and NumPy background and reference some concepts from astronomy, but don't assume any experience with compiled languages (including the CUDA sections, where we stick with Python).
This is the easiest option, unless you want to experiment with the parallelization material in Lecture/Problem Set IX (the free tier of Google Colab only offers 1 physical CPU core).
| Notebook | Link |
|---|---|
| Lecture IX: High-Performance Python: NumPy, Numba, and Parallel Numba |  |
| Problem Set IX: Let's Write a 2PCF Code! | |
| Lecture X: GPU-Accelerated Python with CuPy and Numba | |
| Problem Set X: Practicing GPU-Accelerated Python | |
To run GPU code, you'll need to specifically select the runtime type in Colab:
- Select the arrow next to "Connect" in the top right
- Select "Change runtime type"
- Choose a GPU (the T4 should be plenty for these notebooks)
- Click Save
Flatiron Institute hosts a public BinderHub that allows several CPUs per user, as well as a small GPU slice. This is a good option for working on CPU parallel code, or as an alternative to Colab GPUs. However, the service can only handle a few dozen simultaneous users.
| Environment | Link |
|---|---|
| CPU-only |  |
| GPU-enabled | |
See here for general FI BinderHub documentation: https://wiki.flatironinstitute.org/Public/UsingFiBinder
You can also run any of these notebooks locally. You might want to follow these steps to set up a conda environment:
$ git clone https://github.com/lgarrison/lsst-dsfp-session21
$ cd lsst-dsfp-session21
$ conda env create -f environment.yml -n dsfp-21For Lecture and Problem Set X, you'll need an NVIDIA GPU to run the CUDA code. You can comment out all the cu* dependencies in environment.yml before creating the environment if you're not planning on using CUDA.