Couplage de modèles, algorithmes multi-échelles et calcul hybride

Abstract : In this work, we investigate the implementation of hybrid methods on heterogeneous computers in order to achieve numerical simulations of multi-scale problems. The hybrid numerical method consists of coupling methods of different natures to solve the physical and numerical characteristics of the problem. It is based on a remeshed particle method that combines the advantages of Lagrangian and Eulerian methods. Particles are pushed by local velocities and remeshed at every time-step on a grid using high order interpolation formulas. This forward semi-lagrangian method takes advantage of the regular mesh on which particles are reinitialized but is not limited by CFL conditions.We derive a class of high order methods for which we are able to prove convergence results under the sole stability constraint that particle trajectories do not intersect.In the context of high performance computing, a strong portability constraint is applied to the code development in order to handle the rapid evolution of architectures and their heterogeneous nature. An analysis of the numerical efficiency of the GPU implementation of the method is performed and extended to multi-GPU platforms. The hybrid method is applied to the simulation of the transport of a passive scalar in a 3D turbulent flow. The two sub-problems of the flow and the scalar calculations are solved simultaneously on multi-CPU and multi-GPU architectures.
Document type :
Theses
Complete list of metadatas

Cited literature [123 references]  Display  Hide  Download

https://tel.archives-ouvertes.fr/tel-01094645
Contributor : Abes Star <>
Submitted on : Thursday, June 1, 2017 - 2:36:16 PM
Last modification on : Saturday, February 23, 2019 - 1:22:10 AM
Long-term archiving on : Wednesday, September 6, 2017 - 6:43:12 PM

File

ETANCELIN_2014_archivage.pdf
Version validated by the jury (STAR)

Identifiers

  • HAL Id : tel-01094645, version 3

Collections

Citation

Jean-Matthieu Etancelin. Couplage de modèles, algorithmes multi-échelles et calcul hybride. Analyse numérique [math.NA]. Université de Grenoble, 2014. Français. ⟨NNT : 2014GRENM072⟩. ⟨tel-01094645v3⟩

Share

Metrics

Record views

428

Files downloads

709