ParaDiGM: a library to handle Parallel Distributed General Meshes - Inria - Institut national de recherche en sciences et technologies du numérique Accéder directement au contenu
Autre Publication Année : 2022

ParaDiGM: a library to handle Parallel Distributed General Meshes

Résumé

During the last decade, the Moore's law has proven to be outdated. The computational power does not double anymore every year for a given cost under the pursuit of the energy efficiency. To address this issue, the High Performance Computing (HPC) industry showed an incredible ingenuity by designing complex architectures. Today, the HPC tends to have more and more cores on each node, possibly with accelerators like General-Purpose Graphics Processing Unit (GPGPU) or Field-Programmable Gate Array (FPGA), and with less and less memory. As a corollary, the developers of scientific computing software need to pay attention and master numerous key concepts, which require great expertise, to hope to extract a reasonable performance. These key concepts include programming on heterogeneous architectures, multithreading programming, Message Passage Interface (MPI) programming, load balancing, asynchronous and non-blocking communications, cache blocking techniques, vectorization or even parallel Input/Output (I/O). It is in this context that the ParaDiGM library was created to provide the developers a progressive framework, which consists of a set of low-, mid- and high-level services usable by many developers of scientific computing software that rely on a mesh, typically when solving partial differential equations on a given domain. The ParaDiGM kernel is based on the key concept of a distributed or partitioned view of data stored in an array. After introducing ParaDiGM and giving some general information in the first section, we shall attach great importance in the second section to describe in details the distributed and the partitioned approaches and their impact on parallel algorithms. For instance, we will show that the partitioned view is not always well suited for parallel algorithms and that the distributed view, even though less intuitive, is better suited for specific algorithms. New MPI communication protocols have been developed to switch from one view to the other. Then, we will present parallel dedicated algorithms for wall distance computation in the third section and isosurfaces computations in the fourth section. In each section, some applications will be shown with a special emphasis on the parallel computational efficiency on several thousand cores.
Fichier non déposé

Dates et versions

hal-04401301 , version 1 (17-01-2024)

Identifiants

  • HAL Id : hal-04401301 , version 1

Citer

Eric Quémerais, Bastien Andrieu, Bruno Maugars, Julien Coulet, Stéphanie Lala, et al.. ParaDiGM: a library to handle Parallel Distributed General Meshes. 2022. ⟨hal-04401301⟩
16 Consultations
0 Téléchargements

Partager

Gmail Facebook X LinkedIn More