Parallel finite element solver PARFES for the structural analysis in NUMA architecture. (December 2022)

• Record Type: Journal Article
• Authors: Fialko, Sergiy
• Is Part Of: Advances in engineering software. Volume 174(2022)
• Publication Date: 2022-12
• Subjects: Structural analysis -- Finite element method -- Sparse direct solvers -- Multicore computers -- NUMA architecture -- Multithreading; Computer-aided engineering -- Periodicals; Engineering -- Computer programs -- Periodicals; Engineering -- Software -- Periodicals; Periodicals; 620.0028553
• DOI: 10.1016/j.advengsoft.2022.103290
• Languages: English
• ISSNs: 0965-9978

Abstract

Highlights: The method of multi-threaded parallelization of the assembly and numerical factoring of the symmetrical sparse matrices, realized in the super-nodal finite element solver PARFES and oriented on multi-core computers of NUMA architecture, is proposed. An algorithm for binding RAM to NUMA nodes is presented, which allows minimizing the access of the cores of each NUMA node to the memory of other NUMA nodes (to a distant memory). Parallelization of the sparse matrix assembling uses atomic operations to avoid the collision between threads. It ensures a stable speed-up even in the case of a large number of threads. Parallelization of the sparse matrix factorization is based on the use of a dependency vector that controls the mapping of computational tasks to threads. This approach proved its efficiency on numerous real-life examples of the finite element problems of structural and solid mechanics. Abstract: This paper considers the implementation of PARFES – the direct supernodal method for solving systems of linear equations with symmetric matrices resulting from the finite element method applied to problems of structural and solid mechanics. Unlike previous publications describing the implementation of PARFES for multicore SMP computers, this paper focuses on the implementation of the solver on NUMA computers, where storing data in near memory of the corresponding processor is crucial for achieving high performance, which leads to significant changes in the basic solver algorithms.