Task-based parallelization of a CFD code over a Runtime System

Jean-Marie Couteyen 1, 2 Jean Roman 1 Pierre Brenner 2
1 HiePACS - High-End Parallel Algorithms for Challenging Numerical Simulations
LaBRI - Laboratoire Bordelais de Recherche en Informatique, Inria Bordeaux - Sud-Ouest
Abstract : FLUSEPA is an advanced simulation tool which performs a large panel of aerodynamic studies. It is the unstructured nite-volume solver developed by Airbus Defence & Space company to calculate compressible, multidimensional, unsteady, viscous and reactive flows around bodies in relative motion [2](Figure 1). The numerical strategy in FLUSEPA is designed for highly compressible tlows and to remain accurate when using non-Cartesian grids. The meshing strategy is based on multi-overlapping grid intersection which is conservative and allows to quickly mesh 3D complex geometries. The time integration in FLUSEPA is done using an explicit temporal adaptive method [3]. Instead of using a time step imposed by the slowest cell, cells are grouped inside temporal classes according to their maximum allowed time step. This method allows to perform a fewer number of operations at the expense of some complexity. An iteration of the aerodynamic solver is separated into several sub-iterations. Every cell is not integrated during each sub-iteration but cells which share a face with a slower one are interpolated. At the end of the iteration, each cell has reached the same time. Between iterations, temporal classes can evolve. The current version of FLUSEPA is parallelized using a classical MPI-OpenMP approach and domain decomposition: border faces are duplicated and ghost cells are used. However, the way the time is integrated leads to important time wasted in synchronization despite computation-communication overlapping. The time integration implies an order for the cells to be processed depending on their temporal class: neighbor cells must be at the same time during processing. This locality information is partially lost with the current parallelization. In this paper, we present a way to overcome this issue by working on sub-domains inside each process. We capture the dependencies more precisely through the use of a task-based parallel expression with the help of a well suited runtime system.
Type de document :
Communication dans un congrès
27th International Conference on Parallel Computational Fluid Dynamics, May 2015, Montreal, Canada. 2015
Liste complète des métadonnées

Contributeur : Jean-Marie Couteyen <>
Soumis le : vendredi 15 janvier 2016 - 19:35:25
Dernière modification le : jeudi 11 janvier 2018 - 01:51:38


  • HAL Id : hal-01255450, version 1



Jean-Marie Couteyen, Jean Roman, Pierre Brenner. Task-based parallelization of a CFD code over a Runtime System. 27th International Conference on Parallel Computational Fluid Dynamics, May 2015, Montreal, Canada. 2015. 〈hal-01255450〉



Consultations de la notice


Téléchargements de fichiers