動的領域分割および複数GPUを用いたMPS粒子法の高速化

抄録

In this paper, we propose the dynamic domain decomposition for particle methods, in which the full particles domain is decomposed into several sub-particle domains according to their initial positions. In the dynamic domain decomposition, the exchange of the particle data between neighboring sub-particle domains are effectively reduced, since each sub-domain moves along the particle motions. In addition, when the portion of the boundary particle domain to the full particle domain is small, the accerelation of numerical simulations can be achieved through the parallel computing by using the libraries such as MPIs. In this paper, we apply the method of dynamic domain decomposition to the multi-GPU parallel computing and investigate the efficiency of the method for the numerical accelerations. In fact, we will demonstrate that in dam break simulations, where the balk of water is moved from one side of the wall to another, the acceleration is achieved by reducing the ratio between the volume of the boundary particle domain and that of the full particle domain.