Performance evaluation of various discrete element models implemented on GPU

Abstract

The discrete element method (DEM) is established as a powerful numerical technique to understand and model phenomena of media consisting particles. DEM employing inter-particle contacts is the dominating technology applied for the simulation of the 3D behaviour of grains, powders and particulate solids. The main disadvantage of the DEMis related to computational capabilities that are limited by a huge number of particles and a short time step required in simulations. Naturally, to solve the industrial-scale problems the massively parallel architecture of GPUs and GPGPU computing are the obvious options for significantly increasing computational capabilities. The paper presents performance evaluation of various models of the discrete element method (DEM) implemented in the GPU code. DEM models including computation of the external forces, the normal contact forces, the tangential contact forces with the time history dependent friction and torques are considered for quantitative comparison of the computational costs as well as the bonded particle model. The performance of the developed OpenCL code is evaluated solving applications of granular flows and damage of reinforced concrete. The performance measured on NVIDIA® Tesla™ P100 GPU is compared with that attained by running the same OpenCL code on Intel®Xeon™ E5-2630 CPU with 20 cores. The speedup values, varying from 3.7 to 5.7, are observed for different numbers of particles in spite of intensive usage of advanced vector extensions by OpenCL on CPU. Performed analysis reveals that computation of tangential components of the contact force with time history dependent friction model is the most expensive and increases computing time up to 38.1% of the time required for DEM model evaluating only the normal contact force. Computation of torques is less expensive and adds up to 3.8% of the execution time of the DEM model evaluating only the normal contact force. In case of the bonded particle model, the computing time increases up to 19.4% of the time required for the DEM model of granular flows, assuming the linear dependency of the computing time on the number of particles and applying linear interpolation.