大规模带状线性方程组的分层混合并行求解算法

HIERARCHICAL HYBRID PARALLEL SOLUTION OF LARGE BANDED LINEAR SYSTEM

并行计算过程中,全局通信往往会成为影响算法可扩展性的关键因素。因此,针对大规模带状线性方程组,提出基于MPI/OpenMP多粒度混合编程模型的分层并行算法,将全局通信转化为多次局部通信,解决了全局通信所带来的瓶颈问题,提高了大规模带状线性方程组并行求解算法的可扩展性。

Global communication often plays an important role in the algorithm scalability in parallel computing. In order to solve the large banded linear system, we propose a hierarchical parallel algorithm which is based on multi-granularity MPL/OpenMP hybrid programming model. In this algorithm, we convert the global communication into multiple local communication, this eliminates the bottleneck problem caused by the global communication and improves the scalability of parallel solution of the large banded linear system.