短程力分子模拟在Hadoop上的实现及优化  

Implementation and optimization of short-range molecular dynamics simulation on Hadoop

作　　者：焦善飞[1,2] 何晨[2] 豆育升[2] 唐红[1,2] 

机构地区：[1]重庆邮电大学计算机科学与技术学院,重庆400065 [2]重庆邮电大学高性能计算与应用研究所,重庆400065

出　　处：《计算机工程与科学》2013年第2期26-31,共6页Computer Engineering & Science

基　　金：国家自然科学基金资助项目(21073242)

摘　　要：在Hadoop开源云计算平台上运行分子模拟程序,具有节省软硬件投资、缩短模拟时间等研究意义。然而,该平台并不擅长科学计算类应用中所涉及的快速迭代和子任务间通信。为此,在原子分解法基础上提出了三种解决方案并利用"读写HDFS同步法"实现短程作用力有效的分子动力学模拟的并行算法。在一个Hadoop集群上测试和分析了程序的可扩展性、加速比和各部分耗时情况,结果表明在大规模体系模拟中有较好的效果,最高取得了28倍的加速比。实验证明,Hadoop并行技术在分子模拟中有着较高的经济价值和实用价值。Running molecular dynamics simulation programs on Hadoop has many research meanings such as saving hardware and software investment and shortening simulation time. However, Hadoop is not good at implementing the quick iteration and communication between subtasks in these scientific computing applications. Base on atomdecomposition method, the paper proposes three solutions and implements the parallel algorithm of molecular dynamics simulation by using ＂HDFS read/write synchronization method＂. We tested it on a Hadoop cluster and analyzed expansibility, speedup ratio and costtime of various parts, and then obtained good results as high as 28x speedup in largescale system simulations. The experiment results show molecular dynamics simulation on Hadoop is more economical and practical.

关 键 词：分子动力学 云计算 原子分解法 并行算法 HADOOP平台 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]