Reproducibility of simulations and High Performance Computing

One of the major criteria for the scientificity of a research study is reproducibility. In this talk we will present the main definitions around reproducibility. We will examine to what extent computer science and simulation works are concerned. We will give some examples of applications including parallel stochastic simulations, which are too often presented as non-reproducible. Anyone wishing to produce quality scientific work should pay attention to the numerical reproducibility of his simulation results. Significant differences can be observed in the results of simulations if the practitioner does not apply best practices. We will see that even for parallel stochastic simulations, it is possible to reproduce the same numerical results by implementing a rigorous method tested up to a billion threads. It is possible to check parallel results with their sequential counterpart before scaling, thus gaining confidence in the proposed simulation models. This conference will present best practices able to face even the silent errors impacting top supercomputers, including the Exascale machine arrived this year.

Curriculum Vitae
Professor David Hill is doing his research at the French Centre for National Research (CRNS) in the LIMOS laboratory (UMR 6158). He recently supervised research at CERN in High Performance Computing. He was Vice President of Blaise Pascal University (2008-2012) and also past director of a French Regional Computing Center (CRRI) (2008-2010). Prof Hill was appointed two times deputy director (2005-2007 ; 2018-2021) of the Auvergne Institute of Computer Science ISIMA. He has authored or co-authored more than 250 papers and has also published several scientific books.