多核处理器并行计算模型研究  

Research of Parallel Computing Model based on Multi-core Processor

作　　者：李静梅[1] 张岐[1] 王军锋[1] LI Jingmei,ZHANG Qi,WANG Junfeng（College of Computer Science and Technology,Harbin Engineering University,Harbin 150001,China）

机构地区：[1]哈尔滨工程大学计算机科学与技术学院,哈尔滨150001

出　　处：《智能计算机与应用》2011年第3期9-12,20,共5页Intelligent Computer and Applications

摘　　要：针对并行计算机体系结构中没有通用的计算模型这一问题,分析了一些现有的典型计算模型,在同步性、通信方式、参数方面进行比较,以LogGP模型为基础提出一种改进的mzLogGP模型。利用MPI并行算法对满足节点计算资源非独占、网络存在拥塞条件下的并行程序进行分析与测试,通过增加memory层次化层数和网络拥塞指数这两个参数,计算其计算开销和通信开销,将实测时间与预测时间进行比较,可知随节点数的增加系统误差不断减小,说明该新模型能改善并行应用在多核处理器集群平台上运行的性能,具有较好的可扩展性。As there is not a universal model in parallel computer architecture,the paper analyzes some existing typical models,comparing them in synchronization,communication,parameters,then proposes a new model： mzLogGP based on LogGP.After that,the paper analyzes and tests the parallel process using the MPI parallel algorithm in the environment of non-exclusive nodes on the computing resources and existing network congestion,and calculates the computational overhead and communication overhead by increasing two parameters： network congestion and layers of memory hierarchy.In the end,by comparing the measured time with the predicted time,the paper shows that with the increase in the number of nodes,the system error decreases.The results demonstrate that the new model can improve running performance of parallel application in the platform of multi-core processor cluster,and has better scalability.

关 键 词：并行计算 计算模型 memory层次化 MPI LogGP模型 

分 类 号：TP301[自动化与计算机技术—计算机系统结构;自动化与计算机技术—计算机科学与技术]