Parallel solution of sparse triangular linear systems on multicore platforms

Çuğu, İlke
Many large-scale applications in science and engineering require the solution of sparse linear systems. One well-known approach is to solve these systems by factorizing the coefficient matrix into nonsingular sparse triangular matrices and solving the resulting sparse triangular systems via backward and forward sweep (substitution) operations. This can be considered as a direct solver or it is part of the preconditioning operation in an iterative scheme if incomplete factorization is computed. Often, these sparse triangular systems are the main performance bottleneck due to their inherently sequential nature. With the emergence of multi-core platforms, the interest in solving sparse triangular linear systems effectively in parallel has grown. In this thesis, a parallel sparse triangular linear system solver based on the generalization of Spike algorithm is proposed. The performance constraints of the proposed algorithm and their impacts on the performance are evaluated on matrices from different application domains. Furthermore, performance comparisons are made against the state-of-the-art parallel sparse triangular solver of Intel's Math Kernel Library.
Citation Formats
İ. Çuğu, “Parallel solution of sparse triangular linear systems on multicore platforms,” M.S. - Master of Science, Middle East Technical University, 2018.