Cerfacs Enter the world of high performance ...

🎓Robin CAZALBOU Thesis Defense

  Monday 15 December 2025 at 14h00

  Phd Thesis       Salle JCA, Cerfacs, Toulouse     

Optimizing code coupling in a massively parallel hybrid CPU-GPU environment

MEGEP (MĂ©canique, EnergĂ©tique, GĂ©nie civil & ProcĂ©dĂ©s) – [Subject to defense authorization]

https://youtube.com/live/HN0rXtBs0Vo?feature=share

Multi-physics numerical simulations use specialized computing codes, each based on its own
models and numerical methods. In order to facilitate interaction in the simplest and most modular way possible, it is necessary to use third-party software, known as a code coupler. The CWIPI coupling library, developed jointly by ONERA and CERFACS, enables exchanges to be made between coupled solvers, despite non-coincident meshes, by means of field interpolation. The aim of this thesis is to design and implement multi-process, multi-GPU geometric location algorithms for the field interpolation in the CWIPI library.
The initial stage of the study involved development of multi-GPUs code for dynamic load balancing. This is a key feature, since only a limited area of the physical domain is usually considered during coupling. The core of the study then focused on building distributed octree structures using multiple GPUs and a bottom-up approach. The objective was to accelerate the computation of interpolation coefficients, which is the most time-consuming phase due to recurrent calculations during movements or dynamic remeshing of the coupling zone. The traversal of these trees, as well as the elementary tests of intersection or inclusion of cells, were also developed on the GPU. Finally, a large-scale rotor/stator coupling case was tested to assess the impact of such a coupling on a configuration of industrial interest. The results obtained enabled x100 speed-up factors compared to CPU implementations for load balancing and octree construction. The rotor/stator coupling case doubled the location step speed, paving the way for even larger configurations with complete mesh overlaps.

Jury

François BODINUniversité Rennes 1Reviewer
Guillaume HOUZEAUXBarcelona Supercomputing CenterReviewer
Nahid EMADUniversité Paris-Saclay VersaillesExaminer
Pierre FORTINUniversité de LilleExaminer
Florent DUCHAINECERFACSThesis supervisor
Gabriel STAFFELBACHONERAInvited member
Yvan FOURNIEREDFInvited member

No content defined in the sidebar.