Chinese Journal of Computers   Full Text
Title	An Efficient Parallel Coupling Method for Multiphysics Numerical Simulations
Authors	CHEN Jun MO Ze-Yao
Address	(HPCC &Laboratory of Computational Physics, Institute of Applied Physics and Computational Mathematics, Beijing 100088)
Year	2007
Issue	No.9(1559—1566)
Abstract &   Background	Abstract Complex system is usually comprised of numerous, mutually interacting subsystems and can be numerically simulated by coupling parallel codes individually designed for these physics. The parallel coupling problem, emerged from large scale multiphysics simulations, must deal with different grid representations and different domain decompositions between these parallel codes. This paper presents a practical coupling method for parallel numerical simulation of three-dimensional hydrodynamics coupled with laser propagation, where hydrodynamics code is discrete on Euler grid and laser code on moving grid. Firstly, a temporary state is introduced to split the solving problem into two parts, local interpolation and communication. Then, the coupling relationship between different physical processes is abstracted to a parallel coupling graph, and, according to the dependencies between different field attributes, the parallel coupling graph is simplified. At last, a parallel data redistribution algorithm is provided to accomplish the communication. Parallel numerical experiments and analysis have showed that the method is efficient, and the overall speedup is 50.07 when the parallel application running on 64 processors. The method can be generalized to some other multiphysics applications. keywords parallel computing; multiphysics simulation; parallel data redistribution; directed graph background There is an increasing need for multiphysics simulations in various research and engineering and there is a common feeling that coupling the best codes of each physics process will enable more flexibility and simulation quality to the end user. These codes may be designed by different group and have different grid representation and parallel domain decomposition. How to couple these codes efficiently and across these differences, namely the parallel coupling problem, is the bottleneck in the development of large-scale multiphysics software and becomes the hot spot now. This is a multidisciplinary subject. There have been several efforts to provide support for the coupling of parallel codes, including PAWS, Intercomm, and MCT. More advanced software architectures under development are also trying to do that, such as CCA, in which the interactions between parallel components are standardized. But these studies are mainly in the viewpoint of computer scientist, and assume that the coupling information between codes is provided by users. For simplicity and generalization, each of these toolkits provides a common grid presentation to help fulfill the grid transform between different codes. Users must transform the source grid to the common grid, and also the common grid to the destination grid by himself. It doesn’t satisfy all the grid representations in real applications. This paper attempts to provide the parallel coupling method in real application. This work is supported by the National Natural Science Foundation of China (No.60603051). The project is to solve some critical problems in the coupling the different physics codes in large-scale multiphysics simulations, such as across the differences the grid and domain decomposition and so on. The authors have parallized a lot of real physical applications on structured or unstructured grids for many years, and also design new codes with physical scientists for engineering computations.