基于键涨落模型数值模拟的并行优化

发布时间：2018-03-13 16:27

本文选题：并行计算　切入点：高分子表面吸附　出处：《山东大学》2013年硕士论文　论文类型：学位论文

【摘要】：随着大规模计算的需求不断增长,并行计算技术得到了不断发展,TOP500每年都要公布峰值速度前500强的世界高性能计算机排名。现在主流高性能计算机的体系结构发展趋势,是使用基于共享存储的刀片多核处理器搭建机群系统,山东大学高性能机群便是采用这种架构。在并行计算机发展日新月异的同时,并行计算的实际应用发展速度大大落后于硬件的发展速度,应用程序实测性能远低于计算峰值。因此,充分利用起并行计算机的计算特点,优化提升应用的并行性能,比如对高分子表面吸附的并行应用进行优化,使其在机群上实现并行计算,并且提升并行性能缩短计算周期,成为了并行计算中的一项研究课题。键涨落模型是高分子表面吸附数值模拟的经典运动模型,由于其计算量巨大使得单机PC进行计算模拟的时间不可接受,而MPI实现的并行蒙特卡洛抽样方法可以通过扩展PC(或计算核)的个数来计算不同的样本,最后将数据归约计算,当使用480核计算960样本时,将需要9年完成的串行计算加速至5天完成。这种粒度划分最小为每一个样本为一个计算任务,但当模拟的高分子链分子量比较大时,一个独立样本的计算时间也是相当长的。因此在MPI并行基础之上,可以通过区域分解进一步划分并行粒度,区域分解之后的循环迭代易于使用OpenMP提供的编译制导并行化。相比于MPI使用进程通信,OpenMP基于多线程技术,能更好的发挥刀片结点共享存储的优势。使用OpenMP直接并行化的应用程序,可以初步完成在刀片结点上的并行计算,但高效率的发挥并行性能,需要进一步测试、分析、调优,然后得到最合理的硬件资源使用方案。本文基于高分子表面吸附在高性能机群上的MPI并行编程框架,主要工作分为两部分：首先,研究OpenMP编程技术,实现应用热点模块的并行化；其次,研究OpenMP的优化技术,针对高分子表面吸附应用设计并行优化方案。本文对高分子表面吸附应用的OpenMP程序调优工作,均在四路八核刀片上完成,测试结果表明优化方案能有效地提高实际并行性能。本文采用的软件工程优化化方法,可以为将来机群应用在单个结点上的OpenMP调优提供方法、经验和借鉴。具体而言,本文的主要工作如下： 1.给出MPI并行后的高分子表面吸附在高性能机群上的性能测试和分析,验证满足Gustafson定律； 2.在计算长链并行性能提升达到瓶颈的条件下,本文基于长链分段的方法,使用多线程模拟各段内的键涨落运动,以此来替代长链运动的方法。使用OpenMP并行编程实现并行接口MC_Bond_Fluc,完成在四路八核刀片上的性能测试； 3.设计了高分子表面吸附基于键涨落模型数值模拟的OpenMP优化方案,基于软件优化方法论,使用均衡负载、减少并行开销、合理使用内存、提高Cache命中等手段增量式优化,测试得到了性能最佳的OpenMP优化方式,给出高分子表面吸附的OpenMP并行最优方案。
[Abstract]:With the growing demand of massive computing, parallel computing technology has been developing continuously, TOP500 announced every year ranked high performance computer peak speed of the top 500. Now the mainstream architecture of high performance computer trend is the use of shared storage blade build multi-core cluster based on high performance cluster is Shandong University the architecture of the parallel computer. The development change rapidly at the same time, the development speed of parallel application development speed calculation greatly behind the hardware application, the measured performance is much lower than the calculated peak. Therefore, making full use of computing characteristics of parallel computer, optimized to enhance the performance of parallel applications, such as the parallel application of polymer surface adsorption the optimization of parallel computing in cluster, and improve the performance of parallel shorten computing period, and become A research topic in line calculation.
Bond fluctuation model is a classical model of numerical simulation of adsorption of the polymer surface, because of its large amount of computation makes the single PC was used to simulate the time is not acceptable, and MPI parallel Monte Carlo sampling method can be achieved by extending the PC (or the core) the number of calculation of different samples, and finally the data reduction calculation, when the use of 480 nuclear calculation of 960 samples, and will take 9 years to complete the serial computing speeds up to 5 days to complete. The minimum granularity for each sample as a computing task, but the molecular polymer chain when simulation is large, an independent sample of computing time is quite long. Therefore in MPI parallel basis, can be further divided by parallel granularity domain decomposition, iterative domain decomposition after parallelization using easy to compile guidance provided by OpenMP. Compared with MPI using the process of communication, OpenMP Based on multi threading technology, can better play the advantages of shared memory blade node. Using the OpenMP parallel application programs, can achieve simple parallel computing in blade nodes, but efficient use of parallel performance, the need for further testing, analysis, optimization, and then get the most rational use of hardware resources.
In this paper, the polymer surface adsorption in high performance cluster on the MPI parallel programming framework based on the main work is divided into two parts: first, study OpenMP programming technology and Realization of parallel applications module; secondly, optimization technology of OpenMP, the adsorption of polymer surface with parallel design optimization scheme. The OpenMP program optimization on adsorption the application of polymer surface, was completed in four road eight nuclear blade, test results show that the optimization scheme can effectively improve the parallel performance. This paper uses the software engineering optimization method can provide a method for the OpenMP optimization application on cluster's node, and learn from experience. Specifically, the main work of this paper the following:
1. the performance test and analysis of the polymer surface adsorbed on the high performance cluster after MPI parallel are given to verify the Gustafson law.
2. in the calculation of long chain parallel performance to reach the bottleneck conditions, this method of long chain segment based, multi thread is used to simulate bond fluctuation motion within each segment, in order to replace the method of long chain movement. The use of OpenMP to achieve the parallel interface of MC_Bond_Fluc parallel programming, complete performance testing in the four nuclear Road eight on the blade;
3. the design of the polymer surface adsorption of OpenMP optimization scheme for numerical simulation of bond fluctuation model based on the methodology of software optimization based on the use of load balancing, reduce overhead, reasonable use of memory, improve the Cache hit means of incremental optimization, OpenMP optimization test and got the best performance, given the polymer surface adsorption of OpenMP parallel optimal solution.

【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP338.6

【共引文献】