Analysis of Parallelization Strategies in the context of Hierarchical Matrix Factorizations

H-matrices offer log-linear storage and computations costs, thanks to a controlled accuracy loss. This is the reason why they are specially suitable for Boundary Element Methods (BEM). Task-parallelism strategies are applied to tiled/block algorithms to provide powerful and efficient parallel soluti...

Descripción completa

Detalles Bibliográficos
Autor: Carratalá Sáez, Rocío
Tipo de recurso: tesis doctoral
Estado:Versión publicada
Fecha de publicación:2021
País:España
Institución:CBUC, CESCA
Repositorio:TDR. Tesis Doctorales en Red
OAI Identifier:oai:www.tdx.cat:10803/671577
Acceso en línea:http://hdl.handle.net/10803/671577
http://dx.doi.org/10.6035/14101.2021.429106
Access Level:acceso abierto
Palabra clave:Hierarchical Matrices
Programming models
Parallel computing
Multicore
Boundary Element Methods
Tile H-Matrices
Tecnologies de la informació i les comunicacions (TIC)
68
Descripción
Sumario:H-matrices offer log-linear storage and computations costs, thanks to a controlled accuracy loss. This is the reason why they are specially suitable for Boundary Element Methods (BEM). Task-parallelism strategies are applied to tiled/block algorithms to provide powerful and efficient parallel solutions for multicore architectures. The main objective of this thesis is designing, implementing and evaluating parallel algorithms to operate efficiently with H-matrices in multicore architectures. The first contribution is a study in which we prove that task-parallelism is suitable for operating with H-matrices, while illustrating the difficulties of parallelizing its complex implementations. Afterwards, we explain how the OmpSs-2 programming model helped us avoid the described issues and attain a fair efficiency. Lastly, we explain the creation of the open source library H-Chameleon, based on Tile H-Matrices (a regularized version of H-matrices), which is competitive-with-pure-H-matrices precision and compression ratios, and leverages the benefits of tile algorithms applied to (regular) tiles.