Task-Parallel LU Factorization of Hierarchical Matrices Using OmpSs

Task-parallelism has been exposed as an efficient approach for the solution of dense and sparse linear algebra problems. Hierarchical matrices lie in-between the dense and sparse scenarios and, therefore, it is natural to target this niche of problems via a runtime-based solution that has reported s...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:2017 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW) S. 1148 - 1157
Hauptverfasser: Aliaga, Jose I., Carratala-Saez, Rocio, Kriemann, Ronald, Quintana-Orti, Enrique S.
Format: Tagungsbericht
Sprache:Englisch
Veröffentlicht: IEEE 01.05.2017
Schlagworte:
Electronic mail
Hierarchical matrices (H-matrices)
Indexes
LU factorization
Matrix decomposition
multicore processors
OmpSs
Partitioning algorithms
Programming
Software
Task-Parallelism
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:Task-parallelism has been exposed as an efficient approach for the solution of dense and sparse linear algebra problems. Hierarchical matrices lie in-between the dense and sparse scenarios and, therefore, it is natural to target this niche of problems via a runtime-based solution that has reported successful results in the recent past for related linear algebra problems. Concretely, in this paper we investigate the multithreaded parallelization of the LU factorization of hierarchical matrices using the OmpSs task-parallel programming model and runtime. The focus of our study is in the adoption of an efficient storage layout for this type of matrices, and the analysis of the consequences that this decision exerts on the detection of task dependencies, the programming effort, and the performance of the solution.
DOI:10.1109/IPDPSW.2017.124