python-performance-playground

A performance analysis repository for small Python snippets.

Goal

Ever felt the need to test multiple methods to solve the same, self-contained, easily-explainable Python problem, in order to find the most performant one? Me too, but I'm always too lazy to generalize the one/two lines I wrote, prepare a function and run the experiments over a meaningful set of inputs.

I've collected some of my snippets over a few months, and packed them nicely into a repository, along with an accomodating GitHub action which cares about running the experiments and doing the plots.

Software and hardware

Benchmarks may vary across Python versions, for this reason we provide plots for the latest 3 stable versions:

Python version	Branch
`3.10`	`master`
`3.9`	`python-3.9`
`3.8`	`python-3.8`

Benchmarks run on the GitHub Actions runner ubuntu-latest, updated info about the hardware details of the runner available here.

Content

Directory	Content
`./python/`	General snippets in pure Python
`./numpy/`	Comparisons among multiple equivalent NumPy (or Python) snippets
`./dask/`	Benchmarks of equivalent Dask snippets.

Contributing

Snippets welcome! Just prepare a PR following the standard format in the repository:

Find the right place for your snippet (e.g. numpy, python, dask, ...)
Find an appropriate name for your snippet
Provide an .ipynb files which runs the experiments (please use the annotations @kernel, @data, ... like we did for the other snippets)
Let the bot do its work (after your PR is merged).

Examples

Slice VS List write in NumPy	Pseudo-Hankel matrix

Disclaimer

Benchmarks in this repository are usually very simple, use dummy data and are in general dried of any meaning except for the pursue for performance.
Benchmarks in this repository are assumed to be valid only for common use-cases, while industrial or scientific applications might suffer of asymptotic pathological patterns which deserve a customized treatment.
I do not claim in any way that benchmarks in this repository are enough to provide a full view of the performance of the methods treated. For instance I do not probe memory access or caching with tailored data to stress the computer resources. Again, my aim is to see what happens with common use-cases.

Contributors

Francesco Andreuzzi (CERN, SISSA) -- [email protected]

fandreuz / python-performance-playground Goto Github PK