Acceleration and reduction of memory usage for DEM simulations by implementing spatial blocking and optimization of neighbor particle lists

Abstract

In this work, we propose the improved computational method to reduce the computaional time and the memory usage in GPU-based DEM simulations. Our method has simulteneously achieved performance improvement and reduction of memory usage by improvements of data structure of neighbor particle list. In the parallel algorithms of previous computational process in which one thread correspond to one particle, the warp-divergence frequently occurs because the number of neighbor lists is different. On the other hand, our method can reduce warp-divergence in calculation of particle-particle interaction, and we applied spatial-blocking to DEM calculation that is used in stencil computation to reduce cache miss in reference of neighbor particles. Then, we achieved 1.5x speed up and reduction of 35 ％ memory usage. We can also complete 20 million particles and 200000 time step simulation in 3.3 hours by single Nvidia GTX980 GPU. This result shows the feasibility of low-cost large-scale DEM simulation using our method.