Parallel Multiphysics Coupling: Algorithmic and Computational Performances

Multiphysics problems involve the couplings of different sets of partial differential equations. Partitioned methods consider the individual solutions of each set, which upon iterating, converge to the monolithic solution. The main drawback of partitioned methods is the additional iterative loop, which can be done a la Jacobi (parallel) or a la Gauss-Seidel (sequential). The latter method has worse algorithmic properties than the last method, but makes better use of the computational resources. We will assess both the algorithmic and computational performances of these couplings, in the context of multiphysics surface coupling. To enhance the computational performance of the Gauss-Seidel method, we will introduce an overloading strategy together with an MPI barrier using DLB library. This approach makes the Gauss-Seidel method almost as parallel efficient as the Jacobi method. Our methodology is based on simple performance models, and the solution of multiphysics problems to show the validity of the proposed approach.