I am an ATER (Attaché Temporaire d'Enseignement et de Recherche) since September 2016 at the DALI team from LIRMM laboratory and University of Perpignan Via Domitia .

My thesis was under the direction of Philippe Langlois and Christophe Denis (2013-2016). HPC simulations in various scientific domains suffer from failures of numerical reproducibility because of floating-point arithmetic peculiarities. Different distributions of a parallel computation may yield different numerical results. Numerical reproducibility is a requested feature to facilitate the debug, the validation and the test of industrial or large software. In this thesis, we focus on the openTelemac software that implements finite element simulation for industrial and scientific hydrodynamics. We identify and analyze the sources of this reproducibility failure. We define and implement how to recover numerical reproducibility in two openTelemac modules. We also measure that the running time extra-cost of the reproducible version is reasonable enough in practice.

Some generic key words:
  • Numerical reproducibility
  • Parallel computing
  • Floating-point arithmetic
  • HPC
  • Finite element method
  • Domain decomposition
  • Hydrodynamics simulation