面向组件的通信中间件技术研究

发布时间：2018-05-21 07:42

本文选题：组件式软件系统 + MPI　；参考：《浙江大学》2017年硕士论文

【摘要】：随着并行计算在科学和工程技术领域应用的深入,人们不断研究简单高效的并行计算软件框架。传统的并行计算程序因为模块之间相互耦合,导致程序不易维护和复用。组件式的软件框架具有高度的可维护性和可复用性而成为近年来的研究热点。本文设计了一个软件层次分明、可移植、可扩展的针对流数据计算场景的面向组件的并行软件框架,重点研究支撑软件框架的通信中间件的相关技术。本文首先研究了主流的并行计算模型,从原理、编程模型等角度分析了 MPI、OpenMP和POSIX threads三种并行软件中间件。在此基础上设计了软件框架的分层模型。软件框架总体分为组件层、通信隔离层和通信层,其中通信隔离层又分为发布订阅接口层、发布订阅实现层和流数据处理层。本文还研究了组件间的发布订阅通信模型,完成发布订阅接口层和发布订阅实现层的方案设计,实现了两种基于MPI的发布订阅方案,环形Loop方案和组播方案。本文深入分析了流水线并行的原理并依此设计实现了流数据处理层。最后,本文选用一个典型的信号处理算法来测试软件框架。测试结果显示框架设计达到了预期目标。软件系统无论在双核处理器平台还是在多核服务器平台运行,都能充分发挥硬件的并行计算能力,CPU利用率在90%以上。系统实现了高吞吐率和低时延的要求。且随着平台计算能力的提高,系统指标有大幅度提升。系统的可移植性、可重构性和数据的正确性也得到了验证。
[Abstract]:With the application of parallel computing in the fields of science and engineering, people have been studying the simple and efficient parallel computing software framework. Traditional parallel computing programs are difficult to maintain and reuse because of the mutual coupling of modules. Component-based software framework with high maintainability and reusability has become a research hotspot in recent years. In this paper, a component-oriented parallel software framework for stream data computing scenarios is designed, which is hierarchical, portable and extensible, with emphasis on the related technologies of communication middleware supporting the software framework. This paper first studies the mainstream parallel computing model and analyzes three kinds of parallel software middleware: MPI OpenMP and POSIX threads from the point of view of principle and programming model. On this basis, the layered model of the software framework is designed. The software framework is divided into component layer, communication isolation layer and communication layer. The communication isolation layer is divided into publish / subscribe interface layer, publish / subscribe implementation layer and stream data processing layer. This paper also studies the model of publish / subscribe communication between components, completes the design of publish / subscribe interface layer and publish / subscribe implementation layer, and implements two kinds of publish / subscribe schemes based on MPI, ring Loop scheme and multicast scheme. In this paper, the principle of pipeline parallelism is analyzed and the stream data processing layer is designed and implemented. Finally, a typical signal processing algorithm is used to test the software framework. Test results show that the framework design has achieved the desired goal. The software system can make full use of the parallel computing power of the hardware and the CPU utilization rate is over 90%, both on the dual-core processor platform and on the multi-core server platform. The system achieves the requirements of high throughput and low delay. And with the improvement of platform computing ability, the system index has been greatly improved. The portability, reconfiguration and data correctness of the system are also verified.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP311.52

【相似文献】