QQ客服 官方微博 反馈留言

题名	粒子模拟优化策略及分子动力学模拟软件GROMACS向GPGPU移植的研究
作者	严历
学位类别	硕士
答辩日期	2009-06-05
授予单位	中国科学院过程工程研究所
授予地点	过程工程研究所
导师	郭力
关键词	粒子模拟 并行计算 集群测试 代码优化 通信优化 GPGPU GROMACS
其他题名	Optimization Strategy of Particle Simulation and Porting of MD Simulation Software GROMACS to GPGPU
学位专业	应用化学
中文摘要	粒子方法是一种直观而普适的研究方法，广泛应用于化工、生物、能源等领域。粒子模拟是粒子方法的计算机实现，在实际应用中，粒子模拟对计算机的计算速度和内存容量的需求巨大，需要采用高性能并行计算技术来实现。 中国科学院过程工程研究所多相复杂系统国家重点实验室一直致力于高性能并行粒子模拟的研究，充分利用最新计算机软硬件技术和创新的设计思路，搭建了新的大规模粒子模拟并行集群，使粒子模拟在过程工程领域的应用更加深入。 本论文以本实验室新粒子模拟并行集群的建立为课题背景，结合相关计算机技术与粒子模拟并行计算的特点和需求，进行了三个方面的研究工作： 1. 本论文在新粒子模拟并行集群搭建完成并进行了基本配置的基础上使用Intel MPI Benchmarks、Linpack等工具对集群的计算性能和网络进行了测试，并针对测试中发现的问题对集群进行了调整优化，使集群能够稳定高效地运转。 2. 为充分利用集群硬件的计算能力，本论文对粒子模拟程序的计算和通信两方面的优化进行了研究。首先针对粒子模拟代码特点考察了循环优化技术，并通过编译优化、循环向量化等手段对一个三维宏观拟颗粒模拟程序进行了计算代码的优化，提高了其计算速度。进而分析了影响粒子模拟通信性能的因素，提出了相应优化策略，并通过区域划分方式调整、部署用户层通信协议GAMMA等手段对一个二维宏观拟颗粒并行程序进行了通信优化，提高了其通信性能。 3. 新集群采用了最新的GPGPU计算加速硬件，可提供远高于x86处理器的浮点计算能力。本论文通过对GPGPU编程模式以及通用分子动力学模拟软件GROMACS代码的研究分析，论证了将GROMACS移植到以GPGPU为计算核心的平台上的可行性，并实现了GROMACS中非键力计算函数从x86处理器向GPGPU的移植，获得显著的性能提升，计算加速比达到10倍以上，为GROMACS整体移植的实现奠定了基础。
英文摘要	Particle method is an intuitive and universal research method which is widely used in many domains such as chemistry, chemical engineering, biology and etc. Particle simulation is the implementation of particle method on computer. The demand of particle simulation on computing speed and memory capacity is so huge that high performance parallel computing technology is highly needed. A new large scale particle simulation parallel cluster by fully utilizing latest computer software and hardware technology with 120TFlops (single floating point numbers) peak computing power has been established in State Key Laboratory of Multi-Phase Complex Systems, Institute of Process Engineering, Chinese Academy of Sciences to meet the demands. Based on the establishment of the new particle simulation parallel cluster, this thesis completes research work in three aspects: 1. After the new cluster was established and basic configuration was done, its network and computing performance were tested with testing tools such as Intel MPI Benchmarks, Linpack and etc, then adjusted and optimized to make the cluster run stably and efficiently. 2. In order to fully utilize the computing capability of the new cluster, optimization strategy has been investigated for the particle simulation program in computing code and communication performance. Firstly, the computing code of a 3D macro pseudo-particle simulation program was optimized with methods of compile optimization and loop vectorization and its computing speed was increased. Secondly, based on the analysis of the influence factors for the communication performance of particle simulation, the communication performance of a 2D macro pseudo-particle simulation parallel program was optimized by adjusting area divide manner and deploying user level communication protocol GAMMA and its communication performance was enhanced. 3. To taking the advantage of the latest GPGPU computing accelerate hardware in the new cluster, this thesis also attempts to create GPGPU version of a MD simulation software GROMACS. Using GPGPU as computing core, the non-bonded force computing function in GROMACS has been implemented on GPGPU and achieved the speeds up to 10 that would support the feasibility for porting the entire GROMACS software on GPGPU.
语种	中文
公开日期	2013-09-16
页码	100
内容类型	学位论文
源URL	[http://ir.ipe.ac.cn/handle/122111/1302]
专题	过程工程研究所_研究所（批量导入）
推荐引用方式 GB/T 7714	严历. 粒子模拟优化策略及分子动力学模拟软件GROMACS向GPGPU移植的研究[D]. 过程工程研究所. 中国科学院过程工程研究所. 2009.

个性服务

查看访问统计

相关权益政策

暂无数据

收藏/分享

除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。