Multiple-GPU computing of polycrystalline grain growth simulation using multi-phase-field method

The multi-phase-field method is one of the most powerful numerical simulation methods to study microstructure evolutions in practical polycrystalline materials. However, the computational cost for the multi-phase-field simulation is much higher than the conventional phase-field simulation, since the same number of time-evolution equations with multiple phase-field variables must be solved. In this study, a multiple-GPU computing technique using a programming language CUDA and the MPI library is newly developed to accelerate the multi-phase-field simulation. In order to hide communicational time for CPU-to-CPU and CPU-to-GPU communications, we propose an original overlapping method between computation and communication. Furthermore, we decompose whole computational domain into sub-domains to distribute computational load to multiple GPUs. The multi-phase-field computation for each sub-domain are efficiently performed by using multiple stream executions in CUDA. The three-dimensional grain growth simulation performed using our multiple-GPU computing technique with the overlapping method demonstrates that the communicational time can be hidden completely and good weak and strong scalings are achieved.