Parallel Subdomain-Level DGTD Method With Automatic Load Balancing Scheme With Tetrahedral and Hexahedral Elements
Jiamei Mi, Qiang Ren, Donglin Su
Abstract
In this article, an automatic load balancing method for the parallel subdomain-level discontinuous Galerkin time-domain (DGTD) method has been proposed based on the message passing interface (MPI) library. First, a diagram that exhibits the relationship between the computation load (CPU time) and the DoFs has been extracted from numerical experiments for hexahedral and tetrahedral meshes, respectively. Then, an automatic iteration load balancing scheme has been employed to optimize the positions of the interfaces in order to minimize the CPU time differences between subdomains with the help of the diagram obtained in the first step. This scheme comes with a one-click solution using a greedy algorithm after an initial division of the original model, which is the only step that requires an external operation. In addition, the whole load balancing scheme is only involved in the preprocessing part of the code without modifying the system matrices iteratively. Consequently, the load balance can be relatively easily and efficiently implemented, which can significantly expedite the simulation of the large/multiscale models. The effectiveness of this load balancing strategy has been demonstrated via numerical cases.