【作者】 谢夏；

【导师】 金海；

【作者基本信息】 华中科技大学 ， 计算机系统结构， 2006， 博士

【摘要】 做为网格平台的一个重要组成部分,网格信息服务对整个网格系统的性能有举足轻重的影响。网格信息服务机制的性能优劣对运行在网格平台上的网格应用的性能也会产生直接的影响。由于网格是实现资源的协同和共享,在实际的研究中,往往很难在一个实验中完全实现各种网格中的动态变化行为。因此,这种现实的局限性给直接研究网格的信息服务机制带来了很大的困难。研究者需要对网格信息服务的性能特点、存在的问题、影响因素等内容进行充分的讨论。网格信息服务机制的特点分析是实现网格信息服务机制评估的第一步。在分析了影响网格信息服务性能的关键技术与挑战的基础上,比较四种已有的网格信息服务机制,设计了网格信息服务模拟模型,定义关键评估指标。这些指标包括平均响应时间、平均吞吐率、成功的查询响应次数、资源利用率等。在分析了已有信息服务服务机制和模拟工具的基础上,网格模拟器JFREESIM用来对网格信息服务建模和模拟。JFREESIM是一种基于多任务、多调度、多服务资源架构的网格环境模拟工具。JFREESIM使用离散事件模拟引擎SimJava,包括用户界面、实体层、离散事件基础框架和虚拟资源四层。其中,实体层实现了对网格系统中各个功能部件的模拟。用户可以将新的网格信息服务机制的模型在网格模拟器中进行验证和评估,统计和分析测试结果并帮助用户进行改进。对于内容分发网络、分布式网络游戏、高速科学计算等时间敏感型应用,资源提供者与用户之间需要进行大量的数据交互,同时又对实时性有非常高的要求。网络时延和网络跳数是影响执行时间的重要因素。基于网络时延的网格信息服务模型由索引服务、信息提供者、资源提供者和用户组成。用户根据网络时延,按照自己的要求,选择合适的资源,通过调度器完成任务的实现。基于网络跳数的网格信息服务机制计算用户与所需资源之间的网络距离,清除恶意节点,以分布式的方式支持大范围、多属性的查找。通过将两种网格信息服务机制在网格模拟器JFREESIM中进行测试,测试结果表明它们能满足时间敏感型网格应用的需要。可靠性是网格信息服务的另一个重要评估指标。网络通讯、网络的可靠性、资源的可用性等因素是保证资源可用性的网格信息服务机制必须要重点考虑的问题。它将提交资源更新信息的周期作为采样时刻,以成功提交的资源更新信息作为资源可用的判别标准,保守预测资源节点在下一个采样时刻的可用性。提出了成功接收资源信息的统计更新算法、接收失败资源信息的定时更新算法以及保守计算资源可用性算法。将这种网格信息服务机制在网格模拟器JFREESIM中进行测试,测试结果表明能满足要求可靠性保证的网格应用的需要。更多还原

【Abstract】 A grid information service is an important part of a grid platform, and it influences the performance of the grid platform directly. Since a grid supports resource cooperation and sharing, it is very hard to simulate the pervasive resources and dynamic diversity of a grid in an experiment. That is, it is very difficult to evaluate the performance of grid information service. Researchers need to be able to discuss the performance characteristics of a grid information service, potential problems, interactions with application and infrastructure, trace data and so on.Analysing the characteristics of a grid information service is the first step to evaluating the grid information service. This thesis analyzes key technologies and the challenge of evaluating grid information services. A grid information simulationed modeling is then built based on four popular grid information mechanisms. Performance metrics includes: average response time in seconds; average throughput per second; total number of successful query responses; CPU_Load; resource_usage, and successful_requests.We used a popular tool for grid simulation, called JFREESIM, to model and simulate a grid information service. JFREESIM is a multiple tasks, multiple schedulers and multiple resources (MTMSMR) model, which is composed of a user interface, entities, a discrete-event infrastructure, and virtual resources. The embedded discrete-event simulation engine is SimJava. Entities are used to simulate the real components of a grid system. User can test a new grid information service model in JFREESIM to obtain test results and optimize the model.Time is a sensitive factor for some applications, such as content distribution networks, distributed network games, and high speed computing. Mass data is exchanged between providers and users, with a strict limitation on delays. Network latency and network hops are key factors that affect tasks. A grid information service based on network latency includes: an index service, information providers, resource providers and users. Network coordinates are introduced to locate each grid node. Users choose proper resources based on network latency, and the task implementation is completed by scheduling. A grid information service based on network hops is discussed to compute the network distance between a user and candidate resources, detect bad nodes, and support large-scale and multi-attrribute queries in a distributed manner. Both of these were evaluated in JFREESIM, and proved to be suitable for time-sensitive grid applications.Reliability is another important evaluation metric for grid information services. For grid information services based on resource availability, there are some important factors that need to be resolved, such as: network communication, network reliability, and resource availability. This mechanism regards the resource updating frequency as sample time, and uses the successful resource updating information to predict resource availability pessimistically for the next time. Three relative algorithms were addressed in this part. Experiments with JFREESIM showed that they are suitable for grid applications with reliability constraints.更多还原