Augier et al. (2021) acknowledge that ecological impact of computing is an important problem, but object to criticism drawn against use of Python in a comment by Portegies Zwart in Nature Astronomy who estimated CO2 production in a benchmark N-body problem implemented in various programming languages. According to Zwart, the algorithm required an lengthy "loop" with the order of N^2 operations where Python performed poorly. Therefore, the author suggested to avoid Python in scientific codes.
In order to prove their point, Augier et al. (2021) chose to do a few things differently:
- Use optimized compilers along with minimal change to the same Python benchmark code. This leads to remarkable improvement in performance, comparable to FORTRAN and C++.
- Include an implementation in Julia, which also performs well.
- Use super-computing clusters equipped with wattmeters to accurately measure the energy consumption. We see that use of parallelism does not severely increase CO2 production; rather a mild decrease was observed due to reduced time to solution.
In conclusion, their take-home messages from the study are:
- Rewriting existing high-level codes to other programming languages for simply the sake of performance would be counter-productive.
- Programming languages are not 'interpreted' or 'compiled', only specific implementations are.
- Some basic thumb rules: Premature optimization is the root of all evil. Measure, don’t guess.
- Python can be made multi-threaded, ahead-of-time (AOT) or just-in-time (JIT) compiled with Transonic, Pythran, Numba and PyPy.
- Better knowledge of these tool-chains would lead to less missed optimizations
Read the full article:
P. Augier, C. F. Bolz-Tereick, S. Guelton, and A. V. Mohanan, “Reducing the ecological impact of computing through education and Python compilers,” Nature Astronomy, vol. 5, no. 4, Art. no. 4, Apr. 2021, doi: 10.1038/s41550-021-01342-y.
