Instructor’s guide

Why teach this lesson

Python has traditionally not been used in high performance computing environments due to its inherent performance bottlenecks - i.e. being an interpreted language and having an in-built mechanism (the global interpreter lock) which prohibits many forms of parallelisation. At the same time, Python is enormously popular across scientific disciplines and application domains in both academia in industry, and there are by now a number of mature libraries for accelerating and parallelising Python code. Although better performance and parallel scalability will invariably be obtained by writing code in traditional HPC languages (C and Fortran), teaching researchers and engineers to write performant, parallel and GPU-ported code in a high-level language like Python can save substantial development time, make HPC resources more accessible to a wider range of researchers, and lead to better overall utilisation of available HPC computing power.

Intended learning outcomes

By the end of a workshop covering this lesson, learners should:

• Have a good overview of available tools and libraries for improving performance in Python

• Know what libraries are available for efficiently storing, reading and writing large data

• Be comfortable working with NumPy arrays and Pandas dataframes

• Be able to explain why Python code is often slow

• Understand the concept of vectorisation

• Understand the importance of measuring performance and profiling code before optimizing

• Be able to describe the difference between “embarrasing”, shared-memory and distributed-memory parallelism

• Know the basics of parallel workflows, multiprocessing, multithreading and MPI

• Understand pre-compilation and know basic usage of Numba and Cython

• Have a mental model of how Dask achieves parallelism

• Remember key hardware differences between CPUs and GPUs

• Be able to create simple GPU kernels with Numba

Schedule for 3 half-day workshop

Day 1

Time	Episode
09:00 - 09:20	Motivation
09:20 - 10:00	Scientific data
10:00 - 10:20	Break
10:20 - 11:00	Efficient array computing
11:00 - 11:20	Break
11:20 - 12:00	Efficient array computing

Day 2:

Time	Episode
09:00 - 09:40	Parallel computing
09:40 - 09:50	Break
09:50 - 10:20	Parallel computing
10:20 - 10:40	Break
10:40 - 11:20	Profiling and optimizing
11:20 - 11:30	Break
11:30 - 12:00	Profiling and optimizing

Day 3:

Time	Episode
09:00 - 09:40	Performance boosting
09:40 - 09:50	Break
09:50 - 10:20	Performance boosting
10:20 - 10:40	Break
10:40 - 11:20	Dask for scalable analytics
11:20 - 11:30	Break
11:30 - 12:00	Dask for scalable analytics