Studies On Parallel Computing Based On MPI+CUDA For Discrete Element Computation

In order to improve the computational efficiency,the computing process is decomposed into several small parts then run with parallel.During the practical application,the main purpose of parallel computing technology is used to improve the speed and scale in solving problem.Particle discrete element method is used to simulate the physical representation of the discontinuous media objects.By simulating,the interaction and movement between each individual are researched.The results can quickly and easily guide the actual industrial and agricultural production activities,and then greatly improve the productivity of society.In order to obtain a better simulation results,it requires a very large amount of computation in the practical application process.Therefore,it is a research hotspot to how we improve the discrete element,simulation speed.Parallel computing is one of the most effective ways to solve the problem.The purpose of the study is to improve the efficiency of serial processing algorithm which is used to simulate the forces on particles movement in a container,under a parallel environment,The traditional serial program has been used to achieve some of the discrete element method simulation case basis,but due to computer hardware limitations,its execution speed is not satisfactory and the scale mock objects is also very limited.The parallel algorithm in multi-core systems discrete element was researched based on the analysis of existing procedures.At first,a serial algorithm which can realize the corresponding function was analyzed,and the cycle that spent the most time is found out.The longest cycle in this study included particle movement and traverse grid.Then the longest cycle was processed with parallel.In the study,a single program with multiple data streams method was used to solve the problem.During the method,the algorithm strategies were is divided into multiple sub-domains by using the space domain decomposition.According the positions of particles in discrete element program change frequently,the MPI algorithm and MPI+CUDA algorithm were used.During the algorithms,the communication between two different sub domains was accomplished by the dynamic synchronous.And only the new particles in the boundary region were transmitted in the communication process.As a result,it minimized the amount of information in each communication.During the MPI +CUDA algorithm,the efficiency was improved and the running time was reduced with the use of GPU subdividing the parallel particles further.Finally,the efficiency of various algorithms were quantified and compared.