【作者】 王剑；

【导师】 沈沛意；

【作者基本信息】 西安电子科技大学 ， 计算机软件与理论， 2014， 硕士

【摘要】 随着近年来云计算的快速发展以及多核处理器核心数的快速增长，并行开发技术的应用也越来越普遍。各种并行开发技术如MPI、OpenMP等已经非常成熟的应用于各个领域，并且在Google的MapReduce编程模型的带动下出现了很多基于MapReduce模型的开发框架，如Phoenix、Metis、Hadoop等，但是Phoenix和Metis是基于共享内存架构实现的，不能用于分布式集群，而Hadoop集群部署在嵌入式环境下的执行效率不高。因此目前还没有成熟的基于嵌入式平台的分布式计算框架，虽然MPI+OpenMP方法是一个比较常用的方法，但是节点内使用OpenMP来对并行进行控制增加了编程难度，并造成开发和维护上的开销。为了研究Hadoop在嵌入式处理器平台上的应用缺陷，本文首先构建了基于嵌入式多处理器的云计算平台，在其上实现了并行图像处理方法，进行了效率分析，进而发现了目前Hadoop应用于嵌入式环境中的不足。然后对多核平台下的MapReduce开发框架Phoenix和Metis进行了研究，并基于Tilera36众核平台对其进行性能评测，以此来发现这两个框架的优点和不足。最后在上面所述的研究的基础上，重新设计了一个轻量级的面向嵌入式环境的分布式计算框架MPI+Phoenix，该框架在设计中主要考虑了嵌入式处理器节点之间的异构性，以及如何充分利用节点内部的处理器核心来最大化系统运行效率。最后针对该框架进行了图像并行处理的应用和测试，并对其进行Phoenix标准集测试，从而对该框架的各方面性能进行评估，通过测试表明本论文所提出的框架能够提高数据处理的并行性，同时提升程序的执行效率。更多还原

【Abstract】 With the rapid development of the cloud computing and the multi-core processorsin recent years, the application of the parallel technology becomes more and morepopular. Some parallel development technology such as MPI and OpenMP have beenused maturely in various fields. There are many frameworks based on MapReduce byGoogle, such as Phoenix, Metis and Hadoop. But Phoenix and Metis is on account ofthe shared memory architecture. It can not be used in the distributed cluster. However,the efficiency of Hadoop clusters deployed in the embedded environment is not enough.So there is no mature distributed computing framework based on the embeddedplatform. Although the MPI+OpenMP approach is a more commonly used, but it isdifficult to realize the parallel control using OpenMP. Besides, this approach alsoresulted in the overhead of the development and the maintenance.Firstly, in order to study the imperfection of the Hadoop application, this paperbuilds a cloud computing platform based on the multi embedded processors, the parallelimage processing methods is realized. With the analyze of the efficiency, it can befound that the Hadoop used in embedded environment has some blemish. On the otherhand, the MapReduce framework of Phoenix and Metis based on multi-core platformwere studied, and some performance evaluation was carried out on the Tilera36platform to discover the strengths and weaknesses of the framework.Finally, a lightweight framework for distributed computing MPI+Phoenix isdesigned based on the research described above. The main consideration is theheterogeneity between nodes, and how to take full advantages of the processor tomaximize the efficiency of the system. Then the framework of parallel imageprocessing applications are tested to evaluate the performance of various aspects of theframework as well as to test Phoenix standard test set. The result shows thatMPI+Phoenix model proposed in this paper can improve the efficiency of the dataprocessing and the parallel executing of the programs.更多还原