Parallel Programming and Performance Engineering

Course materials for PPPE — lectures, runnable examples, and activities for a hands-on course on making code use the machine well.

Repository layout

course_materials/    lecture notes and notebooks, numbered in order
examples/            one directory per topic, each with a README and runnable code

Lectures

Class Sessions in Presentation Order

Each number session does not necessarily corerspond to a single day of class. We will consume material at the pace appropriate for class.

Examples

Each directory stands alone: a README.md explaining the idea, source you can compile, and (where it matters) measured numbers.

Single core

Memory

Parallelism

Getting started

Python environment for the notebooks:

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
jupyter lab

The C and C++ examples build with the system compiler and no configuration:

cd examples/memory_hierarchy && make
cd examples/pipeline && ./buildandrun.sh     # builds and runs at -O0 through -O3

Toolchains, only needed for the directories that use them:

• OpenMP — brew install libomp on macOS; already present with gcc on Linux.
• OpenCilk — install from opencilk.org. The cilk/ Makefile expects it at ~/opencilk.
• Java — JDK 17+ and Maven for java_rag_pipeline/.

These examples have been tested on at least two machines (Apple M5, MacOSX using clang 17.0, and AMD Ryzen, Pop!_OS 24.04 LTS using clang 16.?). They will not necessarily run on your machine. You can ask Claude or another tool to adapt them to your machine and that should work. In past versions of the course, I have distributed Docker files that provide a uniform execution environment. I think that is now more difficult and less reliable.

This means that the results you get from examples and activities will vary with software, hardware, etc. They many even vary widely, i.e. not have the same form. This is interesting and something to raise with the instructor and TAs.

A note on the numbers

Performance claims in this repository are measured, not estimated, and measurements are machine-specific. Results were taken on an Apple Silicon Mac and on an AMD Ryzen AI 9 HX 370; where the two disagree in an interesting way, both are reported (files ending in .ryzen.md hold the second machine's run). Expect your own numbers to differ. Reproducing a measurement on your hardware and explaining why it came out differently is a large part of the point.

AI policy

The course is designed around working with an AI partner, and much of the prose in these materials was drafted with Claude and edited by Randal Burns; individual files say so where it applies. Use any tool you want unless a specific assignment says otherwise. Exams and quizzes ask for written answers and pseudo-code in your own words. See the syllabus for the full policy.