A reconfigurable parallel algorithm for sparse Cholesky factorization

Abstract

This paper describes an efficient multi-phase parallel algorithm for sparse Cholesky factorization. The algorithm is simple in its concept and takes ideas from Kumar and Gupta [13] and Roman [18]. We adapt the sub-tree to sub-cube mapping strategy introduced by George et al [9] to reconfigurable parallel machines which allows an improvement in communication performances. In the case of regular grid problems our algorithm incurs less communication overhead and is more scalable that the known parallel sparse Cholesky factorization [9, 13]. Furthermore, we extend our algorithm to the case of the block Cholesky factorization. The different simulation results confirm our analysis and produce good speedup.