Performance of parallel solution of a block-tridiagonal linear system on Fujitsu VPP500

摘要

We study the performance of parallel solution of a block-tridiagonal linear system adopting the block cyclic reduction (BCR) method. We develop a parallel code on Fujitsu VPP500 to solve a block--tridiagonal (BT) linear system and measure the speed of the parallel code on VPP500/28 at RIKEN in order to explore the ability to solve a large system. We study the dependence of the performance on the size of the BT matrix and explore the optimum condition to get high speed using our code. We find that our code solves problems with high speed of 31 Gflops and high parallel efficiency by taking advantage of both vectorization and parallelization. The code enables us to solve a larger system of radiative transfer problem in astrophysics than ever before as an application. Numerical results show that the BCR method works very well as a parallel algorithm on parallel computers with a sufficient specification for the communication. Possible performance of the BCR method on a larger system of parallel computers are addressed.