Design And Optimization Of Molecular Dynamics Algorithm For Crystalline Silicon On Many-Core Accelerated Processor

Molecular Dynamics(MD)simulation is a computational method that uses the position and momentum information of particles to obtain macroscopic thermodynamic properties based on the classical mechanics and statistical physics theory,and is widely used in the fields of materials,chemistry,biology and physics,etc.However,molecular dynamics simulation usually confronts many problems,such as large number of atoms,intensive calculation,long iteration time and low simulation efficiency.At present,using the hardware acceleration characteristics of many-core acceleration processor to realize large-scale high-performance simulation of molecular dynamics has become a research hotspot in the field of computational physics,which is of great significance to the home-made of industrial software.Taking the atomic simulation of crystalline silicon system,this thesis realizes high-performance computing of silicon molecular dynamics simulation on two different architecture computing platforms of GPU(Graphics Processing Unit)and DCU(Deep Compute Unit)many-core processors,and carries out a series of effective algorithm design and optimization to resolve the problems of low memory access efficiency,high instruction consumption,insufficient utilization of computing resources in parallel computing,combined with hardware characteristics.The main research contents are as follows,(1)On GPU computing platform and the Songshan supercomputer system,combined with the CPU+GPU and CPU+DCU heterogeneous parallel computing technology and the analysis of the crystalline silicon molecular dynamics simulation principle,the design and parallel acceleration of crystalline silicon molecular dynamics algorithm for two types of many-core processors GPU and DCU computing platforms are realized.(2)The performance bottleneck of the crystalline silicon molecular dynamics algorithm for the GPU computing platform was analyzed deeply,aiming at the problems such as insufficient utilization of equipment resources and memory access in the simulation calculation of the algorithm,and combined with the CUDA programming structure characteristics to optimize the program in parallel.For the problem of low efficiency of data access and memory in the algorithm,the data structure reconstruction and the optimization method of using multi-level memory are proposed to effectively improve the efficiency of multi-threaded parallel computation and multi-level resource utilization.For the problem of high instruction overhead in simulation computation and the conflict of reading and writing the same memory location by multiple threads in many-core processors at the same time,the instructionlevel parallelism of the program is improved by using a combination of loop unrolling and atomic operation.Finally,to address the problem of low computational efficiency due to the existence of data dependency in the simulation computation,the Thrust parallel operation library is used to achieve high-performance parallel computation,which effectively improves the execution efficiency of the algorithm on the GPU platform.The computational performance of the optimized algorithm is 4.76～57.36 and2.00～4.89 times better than that of the international open-source software LAMMPS and HOOMD-blue in the molecular dynamics simulation,respectively.(3)In order to expand the scientific computing ecology of the domestic Songshan supercomputer system,the CPU+GPU heterogeneous crystalline silicon molecular dynamics algorithm was transferred to the platform according to the CPU+DCU heterogeneous programming method.For the problems of data dependence,low memory access efficiency,branch conflict,and high instruction consumption in the algorithm,the optimization methods of Thrust library function,shared memory,branch prediction,and loop unrolling are used to improve the parallel computing utilization of the DCU accelerator,Memory access and instruction-level parallel efficiency.The computational performance of the optimized algorithm is 1.67 ～ 24.65 and 0.70 ～ 2.10 times higher than the simulation performance of the international molecular dynamics open-source software LAMMPS and HOOMD-blue,respectively.In this thesis,the design and optimization of a high-efficiency crystalline silicon molecular dynamics simulation algorithm on GPU and DCU computing platforms are implemented,which effectively improves the computational performance of crystalline silicon molecular dynamics simulation and increases the time and space scale of the simulation.This work will provide an important reference for the algorithm design and optimization of different many-core processors in the molecular dynamics simulation field.