Dynamics simulations of the discrete particle system are mainly restricted by the huge computational cost. For granular systems, new algorithms for graphics processing units (GPUs) accelerated architectures are developed for higher performance. Communications between multiple GPUs are reduced by an asynchronous method between the computation of inner particles and the communication of outer particles. For gas-solid flows, a coarse-grained discrete particle method based on the energy minimization multi-scale model, that is, EMMS-DPM, is developed to massively reduce the computational cost of the particles. To speed up the computation, the codes for the gas phase and the particles were customized for central processing units (CPUs) and GPUs, respectively, and the communication between them was organized asynchronously. To alleviate the load imbalance issue in the parallel simulation of real applications with complex geometries, a two-layer domain decomposition method is developed.
With the above methods, computer virtual experiments on large-scale discrete particle systems are achieved. For example, a 3D full-loop CFB has been achieved by simulating 1.27e11 real particles with 1.27e8 coarse-grained particles at the speed of 1.5e7 particle updates per GPU per second on 135 NVIDIA K80 GPUs. To our knowledge, this is the largest scale and highest performance DPM simulation of a 3D full-loop CFB, in terms of the computational particles used. It is a strong indication that larger and longer simulations can be realized for industrial systems on supercomputers of higher performance in the near future.
Dr Ji Xu got his B. Sc in 2006 from Xi'an Jiaotong University, and then Ph. D in 2012, from Institute of Process Engineering, Chinese Academy of Sciences. He has been associate professor of chemical engineering at Institute of Process Engineering, Chinese Academy of Sciences since 2015. He is interested in developing multi-scale simulation method to realize virtual process engineering for the particulate systems with discrete particle methods, including developments of more efficient and accurate models, and the algorithms for massively large-scale parallel computing on supercomputers of heterogeneous architecture. The developed simulating method can reach to thousands of millions of computational particles on hybrid CPU and GPU supercomputers and long-time simulations of large scale particulate systems are achieved.
