Topological optimization of triangulations on graphics processing units: a bad idea? - Archive ouverte HAL
Communication Dans Un Congrès Année : 2021

Topological optimization of triangulations on graphics processing units: a bad idea?

Résumé

Although it has been around for several decades, mesh adaption is still a very active research topic, especially when coupled to high performance computing. Mesh adaption is an optimization problem where the objective function is a quality measure for triangles. It is solved iteratively by operating local modifications such as node repositioning or edge swap in small patches defined as the triangles containing a given node or edge. The main issue for parallel computing is that patches overlap, and thus while a patch of elements is being optimized, any other patch containing at least one of these elements cannot be optimized at the same time. The most widely used way to solve this issue on both distributed memory and shared-memory architectures is to partition the mesh, and then adapt the mesh within each partition independently while freezing inter-partition boundaries (i.e., nodes or edges shared by two partitions cannot be modified). To ensure the whole mesh gets adapted, it can be repartitioned and re-adapted several times with changing inter-partition boundaries. This coarse-grained strategy is actually used both for mesh generation and adaption. Recently, fine-grained and hybrid strategies have been proposed for mesh generation on shared-memory architectures. The objective of this work is to implement efficiently the topological optimization algorithm of T. Coupez on a Graphics Processing Unit (GPU). While a Central Processing Unit (CPU) can run sequential instructions very fast, it is hoped that with an efficient massively parallel algorithm, more computational power will be accessed on the GPU. In this presentation of things that did not work (as expected), memory latency, occupancy, and other GPU-related notions will be introduced and discussed in order to give insights on what can be done on the GPU and what cannot, especially when finite element meshes are involved.
Fichier non déposé

Dates et versions

hal-03135304 , version 1 (23-02-2021)

Identifiants

  • HAL Id : hal-03135304 , version 1

Citer

Modesar Shakoor, Lambert Delbeke. Topological optimization of triangulations on graphics processing units: a bad idea?. 14th World Congress in Computational Mechanics (WCCM) - ECCOMAS Congress 2020, Jan 2021, Congrès virtuel, France. ⟨hal-03135304⟩
61 Consultations
1 Téléchargements

Partager

More