Simulation Of Phase-field Model Based On MPI+CUDA Parallel

Phase field method is one of the most promising numerical simulation ways in solidification microstructure of microstructure dendrite growth.In order to make the simulation results more close to reality,the grain pure diffusion phase field model of single-phase system that is used to calculate dendrite growth,is coupled with lattice Boltzmann method that calculates momentum,mass and energy transfer,to establish a PF-LBM multi-field coupled model that adapted to simulate the three-dimensional alloy dendrite growth under the action of natural convection.Because of the microstructure numerical model is an intensive computational problem,the simulation time is too long and the simulation scale is too small.To solve these two problems,in this article,a MPI+CUDA hybrid particle heterogeneous parallel computing is used to explore a heterogeneous parallel algorithm with various granularities,which combined the advantage of GPU high efficiency with MPI capacity to enlarge the scale of simulation.Using MPI+CUDA hybrid particle heterogeneous parallel computing to implement the dendrite growth simulation of PF-LBM phase field model,MPI can be used to conduct coarse granularity division,which divides the complete 3d model into single small equal models in different nodes to break through the limitation of simulate scale in single machine.In each node,fine-grained division is implemented by CUDA parallel way to realize the completely parallelized of intra-node,to improve overall computational efficiency.At the same time,"pseudo three dimensional array" programming method is brought up in CUDA programming,which converts the CUDA linear array access to array coordinate access form to make the CUDA programming easier.In compiling ways,by using the Make File file,CUDA functions and MPI functions are written and compiled unified,which reduced the load of writing and compiling MPI+CUDA separately.In CUDA random number generation,different produce methods are presented and their time efficiencies are compared.The problems of random number generation difficult and time consuming of CUDA are solved.Under the same computing environment,MPI+CUDA hybrid granularity heterogeneous parallel computing in the numerical simulation on the PF-LBM model is implemented,and compared with serial computing,MPI and GPU parallel computing.It is concluded that the parallel results consistent with the serial results,inline with the actual experiment.Through the experiment,the following rules are gotten: on the simulation time,with the increase of the number of nodes,MPI parallel method assumes the inverted parabolic of reduced first and increased later.Compare with single GPU parallel,computing time of MPI+CUDA heterogeneous parallel presents the sin function model with increased first and reduced later,then increased again.At the same time,the same node of MPI + CUDA simulation time has been less than the MPI simulation time.On the scale of simulation,with the increase of number of nodes,the simulation scale is keep rising with different parallel simulation methods.Finally,due to the limitations of the simulation environment,the data through a variety of simulation experiments of this article showed: under the condition of computing efficiency close,the largest simulation scale with 21 nodes MPI+CUDA is4203,which is 13 times to single GPU.In the same simulation scale,the speed-up ratio with 21 nodes MPI+CUDA is 57,increased 54% than 21 nodes MPI.Therefore,the MPI+CUDA hybrid granularity heterogeneous parallel in the multi-field coupled model that is put forward in this paper,contains the advantages of GPU high computing efficiency and MPI extendable simulation scale at the same time.