The ever-increasing supercomputer architectural complexity emphasizes the need for high-level parallel programming paradigms. Among such paradigms, task-based programming manages to abstract away much of the architecture complexity while efficiently meeting the performance challenge, even at large scale. Dynamic run-time systems are typically used to execute task-based applications, to schedule computation resource usage and memory allocations. While computation scheduling has been well studied, the dynamic management of memory resource subscription inside such runtimes has however been little explored. This paper studies the cooperation between a task-based distributed application code and a run-time system engine to control the memory subscription levels throughout the execution. We show that the task paradigm allows to control the memory footprint of the application by throttling the task submission flow rate, striking a compromise between the performance benefits of anticipative task submission and the resulting memory consumption. We illustrate the benefits of our contribution on a compressed dense linear algebra distributed application.
The design of scalable parallel simulation codes for complex phenomena is challenging. For simulations that rely on PDE solution on complex 3D geometries, all the components ranging from the initialization phases to the ultimate linear system solution must be efficiently parallelized. In this paper we address the solution of 3D harmonic Maxwell's equations using a parallel geometric full multigrid scheme where only a coarse mesh, tied to geometry, has to be supplied by an external mesh generator. The electromagnetic problem is solved on a finer mesh that respects the discretization rules tied to wavelength. The mesh hierarchy is built in parallel using an automatic mesh refinement technique and parallel matrix-free calculations are performed to process all the finer meshes in the multigrid hierarchy. Only the linear system on the coarse mesh is solved by a parallel sparse direct solver. We illustrate the numerical robustness of the proposed scheme on large 3D complex geometries and assess the parallel scalability of the implementation on large problems with up to 1.3 billion unknowns on a few hundred cores.Key-words: Geometric multigrid, preconditioning, 3D Maxwell/Helmholtz, parallel computing. * Inria, CNRS (LaBRI UMR 5800) and Université de Bordeaux † CEA/CESTA, CS60001, 33116 Le Barp, FranceUn solveur full-multigrille géométrique pour le problème de Maxwell harmonique
Résumé :Le développement de codes de calcul passant réellement à l'échelle pour la simulation de phénomènes complexes est un challenge. Pour des simulations basées sur des EDP définies sur des géométries complexes, toutes les composantes du calcul, qui vont souvent des phases d'initialisation à la phase ultime de résolution d'un système linéaire, doivent être parallélisées efficacement. Dans ce rapport, nous considérons la résolution des équations de Maxwell tridimensionnelles par un schéma multigrille géométrique pour lequel seul le maillage le plus grossier, qui capture correctement la géométrie, est fourni par un générateur de maillages externe au code. Le problème d'électromagnétisme est résolu sur un maillage plus fin qui satisfait les contraintes liant le pas de maillage et la fréquence étudiée. La hiérarchie de maillages est construite en parallèle via une technique de raffinement et des calculs sans matrice sont mis en oeuvre sur l'ensemble des maillages plus fins dans la hiérarchie. Seul le système linéaire défini sur le maillage grossier est construit et résolu par une méthode directe parallèle. Nous illustrons la robustesse du solveur obtenu par la résolution de problèmes définis sur des géométries 3D complexes et démontrons le passage à l'échelle de l'implantation parallèle sur des problèmes de grande taille allant jusqu'à 1.3 milliards d'inconnues.
We consider the electromagnetic scattering problem of an inhomogeneous obstacle. The methodology applied combines a volume finite element method with a boundary integral method. Both are numerically solved in an efficient way and coupled with a domain decomposition method. The main ingredients are: domain decomposition method with Després's transmission conditions and concentric subdomains, Després's integral equations, fast multipole method, and a parallel sparse direct solver. Numerical results on different inhomogeneity and complex geometries are presented.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.