High-performance Python

Speeding up your Python with compilers and parallel execution.

In scientific computing, Python is the most popular programming/scripting language. While known for its high-level features, hundreds of fantastic libraries and ease of use, Python is slow compared to traditional (C, C++, Fortran) and new (Julia, Chapel) compiled languages. In this course we’ll focus on speeding up your Python workflows using a number of different approaches. In Part 1 we will start with traditional vectorization with NumPy, will talk about Python compilers (Numba) and profiling and will cover parallelization. We’ll do a little bit of multithreading (possible via numexpr, despite the global interpreter lock) but will target primarily multiprocessing.

In Part 2 we will study Ray, a unified framework for scaling AI and Python applications. Since this is not a machine learning workshop, we will not touch most of Ray’s AI capabilities, but will focus on its core distributed runtime and data libraries. We will learn several different approaches to parallelizing purely numerical (and therefore CPU-bound) workflows, both with and without reduction. If your code is I/O-bound, you will also benefit from this course, as I/O-bound workflows can be easily processed with Ray.

Date:
Friday, June 7, 2024

Time:
9:30pm–4:30pm with a one-hour lunch break (12:30pm-1:30pm)

Instructor:
Alex Razoumov (Simon Fraser University)

Prerequisites:
Some knowledge of basic Python, numpy, and working with the Slurm scheduler.

Software:
We will provide access to one of our Linux systems. To make use of it, attendees will need a remote secure shell (SSH) client installed on their computer. On Windows we recommend the free Home Edition of MobaXterm. On Mac and Linux computers, SSH is usually pre-installed (try typing ssh in a terminal to make sure it is there).