仪用主从耦合分布式并行处理容错系统体系结构研究

【作者】 刘峰；

【导师】 葛霁光；

【作者基本信息】 浙江大学 ， 生物医学工程及仪器， 2003， 博士

【摘要】 信息社会的物质基础是信息获取、处理、显示、存储传输和交互技术，其中仪器系统是最重要的技术内容之一。现代医学仪器和科学仪器技术是传统仪器技术的继承和发展，它以信息获取、处理和控制为基础达到对客观对象内在的、本质的客观规律、功能和结构的认识，进而人机交互，最终实现有效使用的目的。现代仪器系统是在传统基础上更强调系统信息处理能力。信息获取过程需要实时动态地观测客观对象的多参数和多层次的信息，信息处理过程需要对这些获取到的信息进行高效和高质量的快速实时处理。两者紧密地结合从而保证了现代仪器系统完成信息融合和系统特征建模的任务。为了构建高性能/价格比的现代医学仪器和科学仪器系统，本文对现代仪器用主从耦合分布式并行处理容错系统体系结构进行了较为深入的研究。 本文概述了信息化仪器系统的发展过程、现状和方向，并综述了主从耦合和分布式并行处理系统体系结构的主要研究。 首先论述了系统体系研究和构建的方法学，提出了适用于现代仪器系统设计构建的重要原则和系统化设计构建的生存周期模型。在研究了主从耦合和通用分布式并行处理体系结构的基础上，根据现代仪器信息化系统的重要特征，创新地提出了基于通信体系和容错体系融合的、适度集中的主从耦合分布式并行处理容错系统体系结构。 主从耦合分布式并行处理系统的设计和构造是一项巨大的系统工程，具有投资大、周期长、涉及技术领域广和复杂性高等特点。因此本论文在系统体系研究和构建的方法学的指导下，以Petri网理论和离散事件仿真技术为基础，构建了模块层次式系统性能评价系统，对系统进行层次化的性能建模仿真和瓶颈分析，获得了系统在不同工作负载条件下的性能特征，为仪器系统体系结构的设计和实现策略提供了重要的性能量化指标和指导依据。 通信体系结构的设计与构建是本文研究的重点。主从耦合并行处理系统是具有两个或多个处理单元(Process Element)的集合，它们相互通讯以协同求解一个给定的复杂因果问题的计算建模处理系统。通信体系与处理单元之间的高效融合是解决主从耦合并行处理系统中所有问题的基础。 主从耦合并行系统中处理的并行性提高了系统性能，但同时处理的时空局部性也能提高性能，这是构建主从耦合并行系统过程中需要平衡的重要问题——并行程度。具有更高并行特性的细粒度处理需要处理单元间高性能的同步模型和所提供机制的强大支持。本文提出了高效硬件同步系统方案——全动态栅栏同步模型，并给出了相应的编程原语。 在本文提出的主从耦合和的分布式并行处理体系结构中，适度集中的并行处理节点通过多通道共享总线的拓扑互联形成了高性能的关键处理环节。共享并行总线是处理节点之间高性能通信的重要资源，为了提高资源的利用效率，本文提出了基于时间优先权并具有仲裁事务缓冲机制的仲裁方案。 现代仪器用主从耦合并行处理系统需要更加灵活和更高性能的单元间通信服务作为支撑，为此本文研究并构造了分布式活动信箱(Distributed Active Mailbox Messaging，DAMM)通信子系统。DAMM对主从耦合的并行系统的高性能通信网络功能进行了必要抽象，减少了协议的处理层次，使得通信网络的性能特征可以直接为用户所调用，满足了用户对实时性事件处理的需求。同时DAMM提供了必要的共享核心资源管理(中断、时间、协议和缓存管理等)能力，提高了通信服务原语的抽象层次，方便了应用编程和核心实现。 系统运行可靠性要求主从耦合分布式并行处理系统必须具有可靠的故障容错能力。复杂的主从耦合并行系统产生了主从耦合和适度的分布式容错管理问题，以及现代仪器系统的特点是数据量大和处理模型复杂，这些要求系统在尽量少的冗余资源条件下力求保证自身的浙乞〔口弋学体d匕学t立七仑文可靠性。在理论上要求系统局部故障条件下，避免产生系统整体失效的可能性，所以系统中的处理问题主要集中在监测系统失效和故障事件的机制(硬件和软件、局部和系统)、系统状态阶段性保存、系统状态较完整恢复等方面，这些都是研究重点。其中关键在于主从祸合的自动转换机制。 为此本文提出了层次式(硬件和软件层次)多机制(状态空间监测和超时监测)系统错误和失效监测体系方案，以及在与通信系统融合的基础上提出了适用于系统状态阶段性保存的轻量级和重量级结合的校验点方案和恢复方案。 现代仪器用主从藕合分布式并行处理容错体系结构的研究涉及面广，涉及问题复杂，构造实现难度大，除了研究了仿真系统及其集成问题以便在系统未构成前可以在仿真系统上进行深入和广泛的实验研究之外，并描述了仿真系统的整体结构。本文最后给出了未来研究发展方向。更多还原

【Abstract】 Essential foundation of information society depends on information acquisition, processing, display, storage, communication and interface technique. Instrument system lies on the most important position in information technology. Based on information acquisition, processing and controlling, modern medical instrument and science instrument technology evolving from traditional instrument technology is capable to help us recognizing instinctive and essential action regularity of target object, then interfacing with it, and utilizing it eventually. Modern instrument system focuses much more on the capability of information processing than traditional system. It is essential to real-time detect multi-parameter and multi-level information of target object dynamically for the key function of information acquisition of instrument system, however, to real-time process those acquired information for that of information processing high efficiently and qualitatively. Modern instrument system incorporating those two phases efficiently is to guarantee accomplishing information fusion and system modeling. In order to construct high performance-price-ratio modern medical and science instrument system, this dissertation pay fairly deeper and constructive attention on researching modern instrument-oriented fault-tolerant master-slave coupling distributed parallel processing system architecture.This dissertation summarizes that evolution, state-of-art and orientation of information instrument, and reviews main research work on the master-slave coupling and distributed parallel processing system architecture.At first, this dissertation discuss about methodology of research, design and construction of system architecture, at same time presents the fundamental suitable for designing modern instrument system and evolutional life cycle of designing action systematically. On the basis of researching over master-slave coupling and generic distributed parallel processing architecture, according to that" important characteristic of modern information system, it proposes properly-centralized fault-tolerant master-slave coupling distributed parallel processing system architecture innovatively.It is systematic project for designing and constructing master-slave coupling distributed parallel processing system, have those features of high-investment, long -duration, comprehensive technique and high-complicatioin. Considering those causes mentioned above, complying with methodology of researching and constructing system architecture, based on Petri net theory and discrete-event simulation technique, this dissertation construct modular and layered performance-evaluating system to model system performance and analyze its bottleneck at different level, and then to find performance characteristic under the condition of different system payload. The performance results from the discrete-event simulation system can provide important performance-quantized index and tutorial factors for designing and implementing instrument system architecture.The research on designing and constructing communication system architecture is at the core of this dissertation. Master-slave coupling distributed parallel processing system contains the set of two or several Process Element, which communicating with eachother to solve a specific complicated consequent problem over computing, modeling and processing. The communication architecture cooperating with actions of Process Elements efficiently comes into being fundament for solving all problems in master-slave coupling distributed parallel processing system.Parallelism of processing in master-slave coupling parallel processing system helps gain its performance, locality of processing in time-space do too. Therefore, How to balance between those factors in constructing master-slave coupling parallel processing system brings out a crucial problem, which being called as Degree Of Parallel. The thin granularity processing with greater parallelism is subject to high-performance Inter-PE synchronization m更多还原