High-Performance Parallel Computing for Stiffness Equation of FEM.
Yoshio Miyakawa, Akira Matsuda, Takashi Kato · TRANSACTIONS OF THE JAPAN SOCIETY OF MECHANICAL ENGINEERS Series A · 1996
Large-scale finite element analyses generally require a very long time for computation. Most of that time is for solving the stiffness equation, so it is important that the computation time for that part is reduced in practical use. We discuss how the Cholesky decomposition, which is used to solve the stiffness equation of FEM, is parallelized in order to reduce computation time. We point out that the synchronization with barrier, which is used popularly, results in the loss of efficiency in the case of band matrix decomposition, and then show that another synchronization mechanism, eventcount, has the ability to construct an efficient algorithm for parallel decomposition. As a result, a shared memory parallel computer exhibits near-linear speedup by the algorithm.