多体模拟的并行优化及软件架构关键技术研究

发布时间：2018-06-03 02:58

本文选题：多体模拟 + 异构平台　；参考：《国防科学技术大学》2012年硕士论文

【摘要】：多体问题，又称N-body问题，研究多个质点相互作用的运动规律，是力学和天体物理学的基本问题之一，具有广泛的应用范围，既能用于研究天体物理学中宇宙天体的演化和运动，又能用于研究分子动力学中分子的运动规律和变化情况。多体模拟的计算复杂度较大，如果直接计算多体的两两之间的受力情况，计算复杂度达到O(N2)，即使使用改进的算法，计算复杂度也达到O(NlogN)，随着运算规模的爆炸性增长，传统PC机的计算能力远远无法满足N-Body问题的求解需要，因此需要大型并行计算机来解决该类问题。目前巨型机中大多采用的是CPU-GPU异构体系结构，在异构体系结构中，CPU主要负责控制GPU以及对数据的采集和结果的收集，GPU负责计算，CPU在GPU进行计算时处于空闲状态，这对CPU资源造成了极大的浪费。本文在CPU-GPU异构平台上，面向多体问题在计算机中的模拟过程，对天体物理学和分子动力学应用分别进行GPU加速，并基于CPU-GPU异构体系结构对多体模拟提出一种CPU和GPU协同运算的两级并行策略，充分利用异构系统的计算资源；为确保计算性能最佳，，本文总结了使用GPU进行计算的MPI进程最佳数目和最佳位置的公式；本文提出了根据CPU和GPU的运算性能差异提出动态更新任务分配比例的算法，保证CPU和GPU之间的计算负载平衡。以上工作为今后对多体问题和异构体系结构的研究工作奠定了坚实的基础。
[Abstract]:Multi-body problem, also called N-body problem, is one of the basic problems in mechanics and astrophysics, which can be used to study the evolution and motion of astrophysics. It can also be used to study the movement and changes of molecules in molecular dynamics. The computational complexity of multibody simulation is very high. If we directly calculate the force between the two parts, the computational complexity can reach ON2G, even if the improved algorithm is used, the computational complexity will reach ONlogNnn, which increases with the explosive increase of the operation scale. The computing power of traditional PC is far from satisfying the need of solving N-Body problem, so it needs a large parallel computer to solve this kind of problem. At present, most of the supercomputers use CPU-GPU heterogeneous architecture. In the heterogeneous architecture, the CPU is mainly responsible for controlling GPU and collecting data and results. The CPU is responsible for calculating the idle state of GPU. This is a great waste of CPU resources. In this paper, on the CPU-GPU heterogeneous platform, the simulation process of multi-body problem in computer is simulated, and the applications of astrophysics and molecular dynamics are accelerated by GPU, respectively. Based on the CPU-GPU heterogeneous architecture, a two-level parallel strategy of CPU and GPU cooperative operation is proposed to make full use of the computing resources of heterogeneous systems, in order to ensure the best computing performance. This paper summarizes the formula of the optimal number and location of MPI processes calculated by GPU, and proposes an algorithm to dynamically update the task allocation ratio according to the difference in the performance of CPU and GPU, so as to ensure the computational load balance between CPU and GPU. The above work lays a solid foundation for the future research on multi-body problem and heterogeneous architecture.
【学位授予单位】：国防科学技术大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP338.6

【相似文献】