Performance of Python runtimes on a non-numeric scientific code

TL;DR

This study evaluates various Python runtimes' performance and compatibility using a non-numeric scientific code that involves complex graph processing, highlighting differences in optimization and execution speed.

Contribution

It provides a comparative analysis of multiple Python runtimes on a real-world non-numeric scientific application, assessing their maturity and performance.

Findings

PyPy showed significant speedup over CPython.

Cython and Numba provided moderate performance improvements.

Compatibility varied across runtimes, affecting usability.

Abstract

The Python library FatGHol FatGHoL used in Murri2012 to reckon the rational homology of the moduli space of Riemann surfaces is an example of a non-numeric scientific code: most of the processing it does is generating graphs (represented by complex Python objects) and computing their isomorphisms (a triple of Python lists; again a nested data structure). These operations are repeated many times over: for example, the spaces and are triangulated by 4'583'322 and 747'664 graphs, respectively. This is an opportunity for every Python runtime to prove its strength in optimization. The purpose of this experiment was to assess the maturity of alternative Python runtimes, in terms of: compatibility with the language as implemented in CPython 2.7, and performance speedup. This paper compares the results and experiences from running FatGHol with different Python runtimes: CPython 2.7.5, PyPy 2.1, Cython 0.19, Numba 0.11, Nuitka 0.4.4 and Falcon.