Two?level parallel strategy for multifrontal sparse Cholesky factorization
Keywords:
sparse algebra, Cholesky factorization, numerical phase, multifrontal method, high performance computing, dynamic parallelization, task?based parallelism.Abstract
In this paper we consider the problem of parallelization of Cholesky factorization numerical phase for sparse symmetric positive definite matrices. A new strategy for parallelization of the multifrontal method for sharedmemory systems is suggested. This strategy combines two approaches to parallelism organization depending on the elimination tree level. At the bottom of the tree, parallel computing of nodes from a priority queue takes place. At the top levels of the tree, nodes are calculated sequentially, employing multithreaded BLAS procedures. Experimental results on the matrices from the University of Florida Sparse Matrix Collection are given. We show that our implementation is commensurable with MUMPS and MKL PARDISO solvers.Downloads
Published
2018-04-07
Issue
Section
INFORMATICS, COMPUTER ENGINEERING AND MANAGEMENT