I am a permanent CEA computer scientist at Maison de la Simulation where I lead the Science of Computing team. My research focuses on programming models for high-performance computing. I am especially interested in the question of separation of concerns between the simulated domain specific aspects and optimization aspects.
I try to concretize my research into production-level software like the DDC, Deisa, or PDI libraries whose development I started, or Kokkos, a project of the Linux Foundation originating from the US DoE. I also try to apply these libraries in production into real codes like the Gysela code for example.
I lead the CExA project that contributes to the Kokkos ecosystem. I also take part in the NumPEx programme, that intends to contribute to the French Exascale software stack. In this programme, I co-lead the Exa-DoST project about IO and data analysis libraries and tools, in the Exa-DI project, I co-lead the work on software packaging, deployment, and CI/CD, and in the Exa-Soft project, I take part in the work GPU programming models.
Multiple positions are funded and available to work with me:
I participate(d) in the following scientific committees:
I am also a reviewer for multiples conferences and journals (Journal of Grid Computing, IPDPS, Europar, ScalA, CBHPC, MAPREDUCE, …)
Advised by Guillaume Latu, I worked on Gysela5D, a semi-Lagrangian gyrokinetic solver for plasma physics in the field of controlled fusion.
I focused on porting Gysela to the Blue Gene/Q architecture. I also worked to modularize the application by applying a software component approach similar to what I proposed during my Ph.D. Thesis.
Generic Support of Composition Operators in Software Component Models, Application to Scientific Computing
Advised by Christian Pérez, I proposed HLCM, a component assembly model in which new composition operators can be introduced. This model is especially well suited for High Performance Computing thanks to the minimum overhead it introduces and to the optimizations it enables. It has been validated with experiments on the Grid'5000 platform.