GMD2.0的建立--排序算法及其他优化
文献类型：学位论文
| 作者 | 李伯杨 |
| 学位类别 | 硕士 |
| 答辩日期 | 2012-05-31 |
| 授予单位 | 中国科学院研究生院 |
| 导师 | 聂峰光 |
| 关键词 | 分子动力学模拟 GPU Hilbert排序 性能优化 |
| 其他题名 | Establishment of GMD 2.0 Version — Sorting Algorithm |
| 学位专业 | 应用化学 |
| 中文摘要 | 分子动力学模拟（Molecular dynamics simulation，MD）通过计算机模拟粒子微观运动来获得体系性质，是一种研究物质微观机理的有效手段。由于计算强度大，目前MD可模拟的时空尺度还不能满足真实物理过程的需要。GPU（Graphics Processing Unit，图形处理单元）作为近年来快速发展的CPU加速设备，为提高MD程序计算性能、扩展MD可模拟的时空尺度提供了新的选择。 合并访问是影响GPU性能发挥的关键因素，GPU硬件的不断改进对数据合并访问的条件逐步放宽，为通过改善数据组织来加速GMD程序提供了可能。Hilbert空间填充曲线能够将空间中邻近点转化为线性相邻排列，本文将其应用于课题组已建立的基于GPU的分子动力学程序GMD程序1.0版本的数据排序，通过尽可能满足合并访问的要求来发挥GPU加速的最佳性能。在此基础上，本文开展了其他性能优化工作，由此建立了GMD 2.0版本。本论文主要工作内容包括： 1. 通过对GMD 1.0性能考察以及与其他分子模拟软件的比较，提出进一步提升GMD程序计算性能的可行方案。为GMD程序建立了一种基于Hilbert空间填充曲线的粒子排序算法。通过重建数据结构，改变成键力计算方法实现了对成键力信息和排除列表的排序，完成的排序算法具有较强的通用性。 2. 对GMD 1.0程序计算性能的进一步加速进行了多方面的探索。主要包括邻居搜索算法的改进、邻居搜索执行判定、线程块大小（block_size）对程序执行速度的影响以及执行参数选取等，一定程度上对程序加速提供了帮助。 3. 加速后的GMD 2.0计算结果与GMD 1.0一致。与GMD 1.0相比，对于L-J势能模拟算例，加速效果为100%~150%；对于聚乙烯结晶过程模拟算例，加速在16%以上；对于反应分子动力学ReaxFF力场，也具有一定加速效果。 |
| 英文摘要 | Molecular dynamics simulation (MD) acquires systems’ macroscopically properties by means of simulating many particles’ microcosmic movement, which is an effective method to understand the behavior of molecular system at atomic scale. But such simulations are still quite limited in size and timescale to meet the spatio-temporal scales of real world applications because of MD is computationally intensive. Owing to the recent advances in the hardware and software architecture, the graphics processing unit (GPU) provides a potential chance for accelerate MD simulation and enlarging spatial and temporal scales in MD. Coalesced global memory read is the most critical phase for GPU performance. With the advances in hardware, the new device improved the qualification of coalesced read, providing possibility of further accelerating GPU program by reorganizing data structure. Hilbert space filling curve can rearrange the order in which particle positions are stored in memory so that neighboring particles are also nearby each other in memory, which has been applied in GMD, a GPU enabled MD program to meet the requirements of coalesced read, so as to maximize the best GPU accelerated performance. Other optimizations have been investigated. Thus GMD 2.0 has been built. The thesis can be summarized as the follow. 1. By performance analysis of GMD 1.0 and having it compared with other MD software, we found a feasible way to further accelerate GMD. We implement a particle sorting program based on Hilbert space filling curve. By reconstructing the coordinate data, data for bonded force calculation and excluded list, with the recoding of bonded force calculation the asorting method has been integrated into GMD 1.0, which is more general by getting rid of some limitations in GMD 1.0. 2. Other possible ways for further acceleration of the program have been investigated including the improvement of neighbor searching algorithm, optimized block size as well as other parameters tuning. 3. The computed result of the accelerated program of GMD 2.0 is identical with that of GMD 1.0. Comparing with the GMD 1.0, the computing performance of GMD 2.0 has been improved up to 150% for simulation of L-J Potential Energy and about 16% for simulation of polyethylene. It also shows acceleration for the ReaxFF, a chemical reactive force field. |
| 语种 | 中文 |
| 公开日期 | 2013-09-25 |
| 源URL | [http://ir.ipe.ac.cn/handle/122111/1807]  |
| 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 | 李伯杨. GMD2.0的建立--排序算法及其他优化[D]. 中国科学院研究生院. 2012. |
入库方式： OAI收割
来源：过程工程研究所
浏览0
下载0
收藏0
其他版本
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。