Research On Method Of Molecular Dynamics Simulation Based On GPU

With the development of Integrated Circuit(IC) technology, Graphics Processing Unit(GPU) is becoming very powerful, the computing capacity and memory bandwidth of GPU are much better than that of the CPU.Comparing with the realization of clusters and supercomputers, the large-scale data-intensive computing tasks, which are completed by GPU as a coprocessor of CPU, have high performance per watt, high performance per square foot and high performance/price ratio.On the other hand, the molecular dynamics simulation is widely used as a power tool of complex condensed matter system in the fields but not limited to physics, chemistry, biology, materials, medicine, but the computing capacity is a bottleneck restricting to the development of their research.Therefore,the research of molecular dynamics simulation with GPU based on Computer Unified Device Architecture (CUDA) has important theoretical significance and pratical significance.This thesis introduces GPU parallel computing technology based on CUDA and molecular dynamics simulation method in detail and gives the CUDA parallel computing model of intermolecular force which takes 90% of simulation time-consuming, and then analysises the parallel computing feature of intermolecular force.The GPU-based algorithm of molecular dynamics simulation is gived and implemented on a PC with Intel Dual-Core 2.93 GHz CPU and GeForce GTS 250 GPU.The result shows that the GPU-based method achieves a high speedup of up to 20 folds.Based on the analysis of insufficient of parallel computing method of intermolecular force, an improved computing model is proposed and validated.The purpose of molecular dynamics simulation is to obtain macroscopic statistical physical quantities, including thermodynamic quantity.This thesis studies the extraction method of thermodynamic quantity, and then gives a CUDA parallel model and algorithm of the extraction of thermodynamic quantity. The result shows that the GPU-based algorithm with a small error has a better speedup which is up to 190 folds.Radial Distribution Function (RDF), which is used to study the ordering of materials and the correlation of electron, is also a common physical quantity. The extraction method of RDF based on GPU is studied, and then the neighbor particle search method is mapped to GPU's thread parallel processing, and the CUDA computing model of neighbor particle search method is proposed. Without losing particle number,GPU brings a speedup of 360 folds.The research of RDF extraction method based on GPU gives a foundation of GPU's application in the field, such as Smoothed Particle Hydrodynamics (SPH) method, N-body problem, life sciences, nanotechnology.