Gabriel WlazłowskiIntroduction
The priorities for W-SLDA Toolkit developers are as follows (from highest to lowest):
- Correctness of computation.
- Performace.
- User-friendly interface.
The highest priority is computation correctness, which is the fundamental issue in computational physics. The next one is performance, which points to programming techniques used for constructing the computational engine (it is written using C+CUDA+MPI).
W-SLDA Toolkit quality
Within W-SLDA Toolkit, we apply the following strategies to ensure a high level of confidence in the computation correctness:
Level 1: Inspection of the results
It is the most standard method (and in many groups, the only one). Namely,
results are correct if they look to be correct.
While it may look like a weak method of testing computation correctness, it works very well in practice. Typically, we know what to expect as the results (more or less), at least at the level of qualitative behavior. Physics is equipped with various methods, experiments, etc., and making extensive comparisons with them leads to identifying discrepancies (if present). From the implementation point of view, the W-SLDA Toolkit delivers extensive support for results analysis.
Actions taken within the W-SLDA Toolkit to support the level 1 quality checks:
- W-data format: conceptually easy data format that allows a variety of tools/languages to be used for data analysis.
- Integration with VisIt: advanced platform for data visualisation & analysis.
- Auxiliary tools & Extensions: we provide various examples of tested codes for data analysis.
Level 2: Internal tests
While Level 1 actions are sufficient to detect most of the errors in results at the qualitative level, they are not sufficient to assure the correctness of the code computation at the quantitative level. Lubarsky's Law says
every program has at least one more bug,
thus, testing is an endless process that must be executed routinely.
Developers of the W-SLDA Toolkit constantly seek cases that can be used as test cases.
Actions taken within the W-SLDA Toolkit to support the level 2 quality checks:
- Testsuite: an automated system of executing test cases that checks the correctness of computation at the quantitative level. The automatic tests are executed routinely after each new commit to the main engine of the code. Presently, more than 300 test cases are implemented.
Level 3: Open-source & Transparency
The review process by external researchers is a crucial aspect of science. It is standard practice in the case of scientific publication. In the context of scientific codes, it is less popular since it requires you to make your code available to others. However, we also observe that open-source/access movement gets stronger with time in computational physics. It is the next-level method that supports the implementation of high-quality code. It steams from Linus's law:
given enough eyeballs, all bugs are shallow.
Actions taken within the W-SLDA Toolkit to support the level 3 quality checks:
- The W-SLDA Toolkit is open-source code.
- Results reproducibility: Together with publications, we deliver reproducibility packs. Using them, a researcher may inspect the computation process from beginning to end. For example, see arXiv:2201.07626.