Title :
Parallel finite element solver for multi-core computers
Author_Institution :
Tadeusz Kosciuszko Cracow Univ. of Technol., Kraków, Poland
Abstract :
The sparse direct parallel finite element solver PARFES, which is based on L·S·LT factoring of symmetric stiffness matrix, where S is the sign diagonal, is developed for analysis of problems of structural and solid mechanics using multi-core desktop computers. The subdivision of sparse matrix into rectangular blocks and the use of procedures from level 3 BLAS leads to high-performance factorization. Comparisons against multi-frontal solvers and PARDISO from the Intel Math Kernel Library when solving real problems from the SCAD Soft1 collection demonstrate that PARFES is a good alternative to the multi-frontal method and, unlike PARDISO, allows solving large-scale problems using computers with limited amount of RAM, owing to virtualization.
Keywords :
finite element analysis; parallel processing; shared memory systems; sparse matrices; Intel math kernel library; L·S·LT factoring; PARDISO; SCAD Soft collection; high-performance factorization; level 3 BLAS; multicore desktop computers; multifrontal solvers; shared-memory computers; solid mechanics; sparse direct parallel finite element solver PARFES; sparse matrix; structural mechanics; symmetric stiffness matrix; Computers; Finite element methods; Mathematical model; Niobium; Random access memory; Sparse matrices; Symmetric matrices;
Conference_Titel :
Computer Science and Information Systems (FedCSIS), 2012 Federated Conference on
Conference_Location :
Wroclaw
Print_ISBN :
978-1-4673-0708-6
Electronic_ISBN :
978-83-60810-51-4