Implementation study of the Parallel Sparse LU Factorization
Abstract
This paper describes an efficient parallel implementation of our previously published LU factorization method for sparse matrices. The main achievement is the usage of a message-passing paradigm instead of the already implemented shared-memory solver. A necessary presentation of the theoretical context is included first, then we briefly show the symbolic factorization steps (build LU elimination forest, postorder traversal, supernode identification) and the numerical factorization steps (Factor and Update steps and algorithm). The experiments were run on an SGI Origin2000 multiprocessor with 64 nodes.