This repository contains the evaluation exercise for candidates interested in the HSF/CERN GSoC project Parallel Processing Improvements in Julia Jet Reconstruction.
- Please get in touch with the mentors of the project to register your interest.
- Read the task instructions below carefully.
- Fork the repository and work on your solution.
- You may set your fork to private, if you wish.
- Invite the mentors to look at your solution by 16 March.
- We will give you some feedback and advice on whether we recommend you to proceed with a proposal for project.
In this repository you will find a Julia script, serial-euclid.jl that
calculates pairwise Euclidean distances between a large number of points.
- Make sure you can setup Julia and run the code.
-
Your first task is to benchmark the initial serial version of the code, using standard Julia tools.
-
Comment on
- how the benchmarking is done and why, with reference to warm-ups, JIT and any other relevant factors;
- the efficiency of this serial version and on any obvious ways to improve it.
-
Now you should implement a parallel version of the code in Julia that can run on multiple CPU cores.
-
Make sure you benchmark the performance, as a function of the number of threads.
-
Please produce a plot of distance-measures-per-second vs. thread count and comment on the results you find.
Your parallel version of the benchmarking code should contain simple instructions for how to reproduce the results (we will fork it and follow your instructions as part of the evaluation).
- Now imagine you now have to port this code to a GPU, using Julia. What would be the key things to pay attention to to ensure the performance is optimal?
It is permitted to use AI to help you in this project, but please do not use coding assistants to generate your solution. Please include a statement saying to what extent you used AI tools when undertaking the exercise.