Calcul à hautes performances : Reproductibilité et répétabilité des résultats numériques et des mesures de performances
Benjamin Antunes · HAL (Le Centre pour la Communication Scientifique Directe) · 2024
Reproducibility is widely recognized as a fundamental principle of scientific research. Currently, the scientific community faces numerous challenges related to reproducibility, often referred to as the "reproducibility crisis." This crisis has affected many scientific disciplines. In this thesis, we examined the factors in scientific practices that could contribute to this lack of reproducibility. Particular attention is given to the pervasive integration of computing in research, which sometimes functions as a black box. This thesis primarily focuses on high-performance computing (HPC), which presents unique reproducibility challenges. We provide a comprehensive state-of-the-art review of these concerns and potential solutions. Additionally, we discuss the crucial role of reproducible research in advancing science and identifying persistent issues in the field of HPC. We discuss the multiple reasons that can lead to a loss of reproducibility when using computing tools present in many areas of research. These include the importance of open science, rigorous documentation, correct application of statistics, scientific culture, software environments, workflows, and software engineering. We also delve into the issues of high-performance scientific computing, where new factors can impact reproducibility. These include problems related to parallel computing, Monte Carlo simulations using pseudo-random number generators, optimization processes, hardware heterogeneity, so-called "silent" errors, and the challenges posed by new paradigms such as quantum computing. We proposed various case studies on reproducibility in high-performance computing and provided insights and recommendations for researchers. We developed an epidemiological model in C++ to easily parallelize large scale individual based simulations of virus spreading, testing parameters were set for Covid 19 and can be changed to tackle other outbreaks like mpox, or other deases that may arise in the future. We studied the statistical quality of stochastic streams based on the initialization of Mersenne Twister, thus providing information on its parallel initialization. We also investigated various reproducibility issues related to the use of pseudo-random number generators in machine learning technologies. During this thesis, we also examined a hardware aspect by questioning the relevance of the systematic activation of simultaneous multi-threading on computing clusters for performance depending on the application profile. Finally, we studied the reliability of quantum machines, a computing paradigm becoming more reliable with the potential to revolutionize high-performance computing in specific areas.