电磁场问题异步并行迭代算法的分析 (1).pdfVIP

摘要 利用数值计算方法求解电磁场问题已经被广泛应用，线性方程组的求解是其中重要的一 部分。随着电磁场问题的日益复杂，线性方程组的规模也变得越来越大，串行算法已经不能 满足科研工作对计算速度的需求。 针对上述情况，本文首先提出了一种基于高斯-赛德尔迭代的异步并行迭代算法。该算法 首先对矩阵分块得到多个不同的线性方程组，然后在不同的处理机中计算不同的线性方程组， 某个处理机计算完毕后将计算结果发送给其它处理机并接收它们当前的数据，然后再利用接 收到的数据求解新的线性方程组。实验结果表明该算法具有更快的计算速度，并将该算法应 用到二维静电场的求解之中，证明了该算法的实用性。 针对于一些收敛性比较差的线性方程组，本文提出了一种基于值域子空间投影法的异步 并行迭代算法。此算法首先对系数矩阵按列分块得到多个不同的投影方程，然后在不同的处 理机计算不同的投影方程，处理机计算完毕后将计算结果发送给其它处理机并接收它们当前 的数据，然后再利用接收到的数据求解新的投影方程。实验结果表明该算法收敛性比较好,计 算速度也比较快，并将该算法应用到电磁散射问题的求解之中，证明了该算法的实用性。 关键词: 并行计算，线性方程组，高斯-赛德尔迭代，值域子空间投影法，异步迭代，电 磁场问题 I 万方数据 Abstract Solving electromagnetic fields with numerical methods is widely used now, and it is very importrant to solve the linear equations in numerical methods. With the electromagnetic fields becoming more and more complex and the increasing scale of linear equations, the serial algorithms can not do work well. In view of this, it proposes an asynchronous parallel iterative algorithm based on the Gauss-Seidel iteration. In this thesis, large-scale linear equations are partitioned into many different linear equations which are calculated in different processors. The processor sends the results to the other processors and receives their latest data after completing the calculation, and then it continues calculating the new linear equation with the data received . The numerical experiments show that this algorithm has higher processing speed. We apply it into the solution of two dimensional electromagnetic fields, which indicates that this new algorithm is practical. The convergence of some linear equations is poor. For this, it proposes an asynchronous parallel algorithm