基于OpenFOAM框架的分子动力学负载均衡关键技术研究

发布时间：2018-03-09 22:01

  本文选题：分子动力学　切入点：负载均衡　出处：《国防科学技术大学》2013年硕士论文　论文类型：学位论文

【摘要】：分子动力学模拟作为一种计算机模拟方法,已广泛应用于物理、化学、生命科学、材料学以及医药等领域。并且,随着人类对科学探索的深入,使用高性能计算机进行大规模分子动力学模拟已成为必然趋势。然而,高性能计算机在带来强大计算能力的同时也遇到了各方面的挑战,如负载均衡、通信、可靠性、功耗等问题,其中负载均衡问题带来的性能损失尤其突出。为了解决该问题,本文提出了面向并行分子动力学模拟的静动态负载均衡方法,并基于Open FOAM框架实现。本文的主要工作和创新点包括以下几个方面:?建立了基于Open FOAM的分子动力学负载均衡框架(第二章)本文深入分析了Open FOAM并行程序框架及其分子动力学的实现方法,发现负载不均衡会严重影响分子动力学的模拟效率。针对该问题,在原有Open FOAM框架的基础上建立了静动态负载均衡机制和相应的数据组织、计算流程、框架组成以及与其他模块的结合方法。本文提出的面向分子动力学模拟的负载均衡框架对其他粒子方法负载均衡机制的设计具有借鉴意义。?提出了面向并行分子动力学模拟的静态负载均衡方法——MDSLB(第三章)本文提出了一种面向大规模分子动力学模拟的静态负载均衡方法——MDSLB。通过对分子动力学中短程力计算特征的深入分析,我们将短程力进一步细分为三类力模型,每类力模型中涉及的计算任务又被拆分为多个细粒度的独立计算单元,称为“cell负载”。这些计算单元为MDSLB方法的实现提供了基本的数据结构。在MDSLB方法中,模拟区域被分为多个子区域,称为“局部空间”。每个局部空间中的cell负载被均匀地分配给各个处理器。该方法通过在程序开始时执行一次,从而保证整个模拟过程中负载的均衡。?提出了面向并行分子动力学模拟的动态负载均衡方法——MDDLB(第四章)本文针对小规模分子动力学模拟提出了动态负载均衡方法——MDDLB。该方法基于Open FOAM的cell结构,设计了负载迁移的基本单元——近邻粒子列表。另外,还引入了阈值检测机制负责监控和衡量每个处理器的负载情况。当负载不均衡的程度超过一定值时,则从负载较重的处理器将负载动态迁移到负载较轻的处理器上。基于该方法,可以有效解决小规模分子动力学模拟中的负载均衡问题。?验证了负载均衡框架及MDSLB方法和MDDLB方法的正确性和有效性(第五章)本文基于Open FOAM设计实现了第二章所提出的分子动力学负载均衡框架,并基于该框架分别实现了第三、四章提出的MDSLB方法和MDDLB方法。基于Tian He-1A子系统使用扩散模型对上述框架和方法进行了测试,实验结果表明,分子动力学负载均衡框架及在其上实现的两个方法是正确有效的。与传统方法相比,MDSLB方法可以获得34%-64%的加速,而MDDLB方法则可以获得更接近理想情况的加速比。
[Abstract]:As a computer simulation method, molecular dynamics simulation has been widely used in the fields of physics, chemistry, life science, materials science and medicine. It has become an inevitable trend to use high performance computer for large-scale molecular dynamics simulation. However, high performance computer not only brings powerful computing power, but also meets various challenges, such as load balancing, communication, reliability, etc. In order to solve this problem, a static and dynamic load balancing method for parallel molecular dynamics simulation is proposed. And based on the implementation of Open FOAM framework. The main work and innovation of this paper include the following aspects:? In this paper, a molecular dynamics load balancing framework based on Open FOAM (Chapter 2) is established. In this paper, the parallel program framework of Open FOAM and its molecular dynamics implementation method are deeply analyzed. It is found that load imbalance will seriously affect the efficiency of molecular dynamics simulation. Based on the original Open FOAM framework, the static and dynamic load balancing mechanism and the corresponding data organization are established. The proposed load balancing framework for molecular dynamics simulation can be used for reference in the design of load balancing mechanism for other particle methods. In this paper, a static load balancing method for parallel molecular dynamics simulation, MDSLB (Chapter 3), is presented. In this paper, a static load balancing method for large-scale molecular dynamics simulation is presented. An in-depth analysis of computational features, We further subdivide the short-range forces into three types of force models, and the computing tasks involved in each force model are split into several fine-grained independent computing units. Called "cell load." these units provide the basic data structure for the implementation of the MDSLB method. In the MDSLB method, the simulation region is divided into multiple subregions, Called "local space". The cell load in each local space is evenly distributed among the processors. This method ensures load balancing throughout the simulation process by executing it once at the beginning of the program. A dynamic load balancing method for parallel molecular dynamics simulation is presented in this paper, which is based on the cell structure of Open FOAM. In addition, a threshold detection mechanism is introduced to monitor and measure the load of each processor. Based on this method, the load balancing problem in small-scale molecular dynamics simulation can be effectively solved. The correctness and validity of the load balancing framework, MDSLB method and MDDLB method are verified (Chapter 5th). The molecular dynamics load balancing framework proposed in chapter 2 is designed and implemented based on Open FOAM, and the third one is implemented based on the framework. The MDSLB method and MDDLB method proposed in the four chapters are tested by using diffusion model based on the Tian He-1A subsystem. The experimental results show that, The molecular dynamics load-balancing framework and the two methods implemented on it are correct and effective. Compared with the traditional method, the MDSLB method can get 34-64% acceleration, while the MDDLB method can obtain a speedup closer to the ideal condition.
【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】：2013
【分类号】：TP38
，

本文编号：1590434