【作者】 孙华志；

【导师】 杨扬；

【作者基本信息】 北京科技大学 ， 控制理论与控制工程， 2008， 博士

【摘要】 网格和P2P计算是当前分布式计算领域的两个研究热点，两者采用的技术不同，但对于分布在网络中的硬件和软件资源进行共享和协同是两者的共同目标。近年来，在研究和应用领域，网格计算和P2P计算出现相互融合的趋势。针对网格和P2P研究领域存在的问题，Ian Foster指出，虽然网格的体系结构较为完善，但对于体系中单节点失效及其所带来的数据、服务失效等问题尚无完整的解决方案；而P2P对于节点动态变化有较好的适应性，却没有良好的系统模型[1]。能否将二者的优势进行融合并给出融合后系统中诸如服务失效，资源定位与查找等关键问题的解决方法是当前分布式计算领域研究的热点和难点。围绕该问题，本文对网格和P2P的融合问题进行了深入的理论和实践研究。研究中以大系统控制论[2]为指导，提出了基于P2P的Grid模型—GriP模型，该模型将二者的优点即良好的体系结构和健壮的节点故障恢复结合在一起；给出了解决网格中节点失效问题的P-SRing方法；在解决了GriP拓扑稳定性地基础上，提出了在资源管理系统（ResourceManagement System，RMS）中基于元数据的资源共享机制及高效资源发现算法。本文的主要内容与创新点包括：（1）提出了一种基于Grid与P2P相结合的集散式网络计算新模型GriP网格系统结构稳定，但系统不能有效地解决单节点失效问题，一旦某节点失效有可能导致整个系统的崩溃，P2P的最大优点是系统可以适应多个节点的动态变化，系统具有较好的灵活性和可扩展性。本文提出了一种将Grid和P2P相结合的系统模型—GriP，将网格技术和P2P技术有机地结合，既弥补了各节点组织不够灵活和节点不稳定的缺陷，又充分发挥了网格和P2P各自在节点控制方面的优点；模型中还提出了基于元数据的资源管理机制。该模型在融合了Grid和P2P各自优点的同时避免了两种模型本身固有的弊端。（2）提出了一种适应单点失效的双层混合型星环式网络拓扑结构P-SRing本文提出了一种“双层混合型”网络拓扑结构。其中，资源节点按照环状拓扑结构分布在低层，在高层中代理节点又与资源节点遵循星型网络的拓扑结构。在环状拓扑结构层中，每一个资源节点都以其邻节点作为其后继节点；某些节点发生故障时，通过本文给出的双层路由机制使整个系统具有更高的安全性与可靠性，当节点发生故障时，其后继节点会自动与相关节点进行连接，使网络能正常、平稳地运行。（3）提出了一种基于资源位置与用户兴趣的GriP节点定位快速算法在网格中进行资源查找分为两个步骤，资源节点快速定位和资源内容查找，针对资源节点的定位，本文提出了基于资源节点位置与节点反馈的资源搜索算法，利用位置信息对节点分群，保存历史查询的搜索反馈结果，计算邻居节点的预计搜索成功性，动态选择转发与实际位置相邻的节点以减少网络开销和发现资源的时间，通过计算邻居节点的兴趣相关度结合实际地理位置，提高了网络资源的搜索稳定性与查询效率。（4）提出了一种基于元数据与平衡树的GriP资源内容查找算法为了提高GriP系统中资源的查找效率，对于资源内容查找我们研究了用于标引资源的元数据的高效搜索算法，GriP中采用了基于XML的元数据标引方法，在此基础上提出了基于关键字的资源查找算法QBVT，对树的结构进行了改进，将父节点的信息增加到子节点中来，通过构建平衡树提高了基于资源内容的查询效率，降低了树的更新开销。更多还原

【Abstract】 At present, Grid and P2P Computing are two hotspots in Distributed Computing researchfield. Grid and P2P Computing apply different technologies, however, they are both aimed atsharing and collaborating hardware and software distributed in network. Recently, there is atrend that Grid and P2P Computing merge in research and practice fields.In solving the existing problems in Grid and P2P Computing fields, Ian Foster pointed outthat Grid had a comparatively perfect structure, but it couldn’t solve problems such as single-node failure and related data and service failure perfectly; while P2P could adapt well to thenode changed dynamically, but it hadn’t good system architecture[1]. Whether we can mergethose advantages and then give solutions to service failure, resource locating and informationsearching becomes a current hotspot and difficulty in distributed computing field.Focusing on this problem, this paper studies intensively the problem of Grid and P2P’smerging with theories and practices. Firstly, Large-Scale Systems Cybernetics[2]is as a guide,and GriP model based on P2P and Grid is put forward. This model combines good architectureand robust node fault recovery; Secondly, it presents P-SRing method, is a solution to nodefailure in grid; Finally, after achieving GriP’s topology stability, resource sharing mechanismand efficient recover algorism in RMS (Resource Management System), which are based onmeta-data, are proposed.The main contents and innovations are as follows:I. Putting forward a distributed network computing model---GriP, based on P2P and Grid.Nodes in Grid are relative stable, but Grid can’t solve single-node failure problemeffectively. Some node failure may leads to the breakdown of the whole system. The mostimportant advantage of P2P is that the system can adapt to multi-node dynamic changes, andhas good flexibility and scalability. Therefore, this paper proposes a GriP model, combined Gridwith P2P. By merging Grid and P2P, we can not only remedy defects of node structureinflexibility and node instability respectively, but also bring their respective advantages of nodecontrolling into full play. This model draws advantages of Grid and P2P and gets rid of theirinherent disadvantages.II. Presenting “Star-Ring-Mixed” P2P topology structure P-SRing which adapts to single-node failure and related problems. This paper presents a “two-layered mix” network topology structure, in which resourcenodes distribute using ring topology at lower layer, while every resource node connects proxynode using star topology at higher layer. At the ring layer, every resource node takes itsneighbor node in clockwise direction as successor node; when some node breaks down, doublerouting mechanism make the whole model more safe and stable. Once some node breaks down,its successor node will connect to related nodes automatically, and the whole network can stillrun normally.Ⅲ. Proposing an algorithm on how to rapidly locate the node in GriP based on resourcelocation and users’ interests.There are two steps for searching for resources in Grid, which are rapidly locating resourcenodes and searching for resource contents. Concerning the first step, this paper proposesresource node search algorithm based on resource location and node feedback. This algorithmgroups the nodes according to location information, saves historical search feedback result,computes neighbor node’s estimated degree of successful searching and dynamically forwardsinformation to physically connected node so as to save bandwidth and shorten time of findingthe first file. By computing neighbor node’s interest correlation degree and adjusting neighbornodes periodically according to geographical position, the resource node search algorithmimproves stability and efficiency in resources searching.Ⅳ. Introducing a resource searching method in GriP based on Metadata and Balance Tree.As for resource content searching, in order to improve search efficiency in GriP, westudied effective search method, which is used to index metadata of resource. GriP appliesXML-based metadata to indexing method, and further puts forward a keyword-based resourcesearch method–QBVT. This method improves the tree structure, which adds parent node’sinformation into child nodes, and constructs a balance tree so as to reduce cost of tree updatingand improve search efficiency based on resource contents.更多还原