【作者】 徐陈锋；

【导师】 奚宏生；

【作者基本信息】 中国科学技术大学 ， 控制理论与控制工程， 2008， 博士

【摘要】 随着网络和信息技术的发展,基于计算机网络的各种应用和创新不断出现。在当前网络技术研究和应用中,P2P(Peer-to-Peer)和IPTV(Intemet Protocol Television)都是人们关注和研究的热点。由于节点或用户的频繁进出,这两类系统都具有极大的随机性。而且在大规模应用中,这些系统在网络中的覆盖面往往极为广泛,并且和其它大量服务集成在一起,表现出分布式分层的结构特征,确定其性能的主要因素往往很复杂,并且各种因素相互制约。应用系统的性能瓶颈问题和优化问题成为研究、设计和实现这些系统的关键问题,其中有些可以直接通过技术改造和升级来解决,但需要大量人力物力。从系统理论的观点来看,其中一些问题也是可以在现有环境下通过现代优化和控制的理论和方法来解决的,特别是,还可以利用基于Markov过程的模型,以避开具有复杂物理背景的精确数学模型的构造与辨识。考虑一类具有分层结构的非结构化P2P系统的层级构成和“组划分”问题。这类系统按照某种原则分成一个个由若干节点构成的组,并在每个组中选出一个超级节点来索引本组内的节点和资源,普通节点的查询通过超级节点的组内查询和组间查询完成。这类系统实际将网络分成了两层,第一层是组内系统,以超级节点为中心;第二层是以超级节点为代表的组间系统,采用纯分布式P2P方式和使用泛洪式搜索。由于系统规模和节点分布是随时间动态变化的,组的划分需要进行适应,以保汪最优的系统总体性能。本文将这类系统的“组划分”及其切换行为进行抽象,利用Markov切换空间模型进行描述:每一个组划分下的状态过程对应于一个状态空间,其动态性可以用Markov过程描述;而组划分切换的控制对应于Markov切换空间模型的行动,根据给定的策略进行选取。可以证明Markov切换空间模型等价为Markov决策过程或参数化Markov报酬过程,在此基础上,利用策略迭代和梯度方法求解最优策略并通过仿真验证所获结果的正确性。在实际工程中的一种具有P2P特征的分布式VoD(Video on Demand)系统,采用了全局中心化的目录服务器对存储到网络边缘服务节点上分段缓存的数据进行索引和定位。针对这种中心化资源定位服务存在的问题,使用将随机漫步和中心目录服务器结合的混合搜索方法提高系统的扩展性和降低目录服务器或网路单点失效问题的影响。在此基础上,利用基于Markov过程的模型来描述和讨论采用这种方法的资源定位服务。为了应对大状态空间的问题,保证状态空间不会进一步扩大,本文利用基于事件的方法进行建模,并根据状态的不同性质进行状态聚集,以降低问题的复杂程度。还利用简化的一次定位模型,讨论了混合方法的搜索效率和代价的控制问题。考虑到实际中接入控制和定位服务是息息相关的,本文将接入控制引入资源定位服务模型,和定位服务结合在一起进行了讨论。更多还原

【Abstract】 With the development of network and information technology, there are more and more various applications and innovations based on computer networks. Recently, P2P (Peer-to-Peer) and IPTV (Internet Protocol Television) are those famous research and application fields of the current network technology. They are both stochastic systems due to the frequent accesses of their nodes or users. And they are generally integrated with many other services and cover a large range of Internet. Therefore, each of them may have a distributed and hierarchical structure and there exist many complicated factors constraining each other in system performances. The performance bottlenecks and optimizations have been become the key problems currently on research, design and implementation of these application systems. Some of them can be solved by direct technical reconstruction or update, which is generally not low-cost. From the view of system theory, some of them can also be solved by using modern optimization and control theories and methods. Especially, the models based on Markov processes can also be applied in order to avoid constructing and identifying the accuracy mathematical descriptions with complicated physical backgrounds.Considering the hierarchical structure and switching problem for a class of hierarchical unstructured P2P systems. According to some given principle, these systems are divided to many groups with lots of nodes and elect a supper node for one group. Each supper node indexes the nodes and resources in its own group. A normal node has to search resources via the super node’s location service, which consists of the internal search engine in group and the one between groups. These systems actual have a two-layered structure: one is the single group around its supper node, the other is the subsystem between groups, which is a pure P2P and uses flooding-like search. Since the system scale and distribution of nodes varies dynamically, the partitioning has to be adaptive to maintain the system’s overall performance. In this paper, a Markov switching-space model is introduced to describe their behavior of dynamic grouping. The state process under each partitioning is characterized as a Markov process with a special state space. The controls of grouping behavior correspond to the actions of Markov switching-space model, which are selected according to some given policy. It can be proved that the Markov switching-space model is equivalent to a Markov decision process or a parameterized Markov reward process. Therefore, the optimal policies for dynamic grouping can be given by applying policy iterations and gradient methods, which are verified by some simulations in this paper.A kind of P2P-enhanced distributed VoD (Video on Demand) system in a practical project has a global centralized directory server to index all the data of segment caches on edge service nodes and to provide location service for them. This solution has poor scalability and suffers from single point of failure problem. Therefore, a hybrid search which combines random walk and the existent global service is proposed. And models based on Markov processes are introduced to characterize and discuss this hybrid location service. In order not to extend the large state space, event is used in models. And also state clustering is applied to reduce computational complexity. Then, a simplified single location model is placed to discuss its effect, search efficiency and control problem of search costs. Considering the close relation between admission control and location service, a more complicated model combined with admission control is also discussed.更多还原