【作者】 陈海勇；

【导师】 赵荣彩；

【作者基本信息】 解放军信息工程大学 ， 计算机软件与理论， 2009， 博士

【摘要】 计算机硬件和网络技术的迅速发展，使得集成广域范围内的计算资源进行分布式计算成为可能。由此自上世纪90年代以来，网格(Gid)技术在世界范围内受到了众多研究机构和商业企业的关注。网格通过聚合异构、分布、自治的各类计算资源，形成一个共享协同的计算环境。网格作业调度是网格系统的重要功能，由于网格环境的广域特征，以及网格结点的动态性、自治性、不可靠性等固有属性的存在，使得网格作业调度成为制约网格技术发展的关键因素之一，是网格相关研究的一个重点和难点。网格经济模型通过将现实中的市场经济基本原理引入网格资源管理与作业调度之中，促使网格平台资源提供方与使用方在利益驱动下构建大规模网格环境，是目前网格及其网格作业调度的研究热点。本文在前期网格经济模型相关研究的基础上，首先对网格的发展、结构、组成及其作业资源分配等相关方面进行了分忻总结，然后针对网格作业调度过程中作业类型支持、运行风险、价格制定策略及调度模拟等若干关键问题进行了进一步研究。本文的主要贡献与刨新总结如下：1、提出了截止时刻与计算预算约束下的系统级网格作业调度方法，改进了传统的直用级网格作业调度只针对参数扫描直用进行优化调度的不足，在兼顾考虑作业初始优先级与优先级动态调整的同时，实现了对传统并行作业网格调度的支持，并为后续工作奠定了基础。2、提出了网格作业调度过程中基于价格的风险控制模型，通过引入风险效用曲线，解决了在网格这种非可靠的计算环境中如何刻画网格用户对风险的承受能力问题，实现了平衡网格结点可靠性、资源价格与用户风险承受能力的网格作业调度算法。3、提出了单个网格结点的资源阶梯定价模型，通过对网格用户进行级别划分和阶梯收费，解决了单一资源定价模型中无法区分不同类型网格用户的问题，实现了通过设置不同的资源规模保障不同级别用户服务质量，提高网格结点的赢利。4、提出了在多级阶梯价格中，直用边界期望收益确定各级别用户最优资源规模的方法，通过建立各级别用户对资源的需求分布函数，解决了多级阶梯价格中各级别用户资源规模不易确定的问题，从而保证网格结点赢利期望的最大化。5、针对网格模拟工具GidSim的不足对其进行了功能扩展，改进了GidSim只能模拟单cPu作业、无法描述作业串行部分、不支持第三方数据传输以及缺乏多cPu作业支持下的资源分配模型的不足，实现了对这些最基本网格特征的功能支持，扩展了GidSim的直用范围和使用灵活性，同时为本文的相关模拟测试提供了适合的实验平台。本文直用改进后的网格模拟工具对本文研究内容进行测试，测试结果表明，通过直用相关算法与模型，能够较好地在网格经济环境中对网格作业进行调度，在满足作业服务质量要求的同时提高网格结点赢利水平。本文通过直用网格经济模型，在系统级作业调度、作业风险控制、阶梯价格模型及相直资源规模调整方面做了较为深入细致的研究，但是仍存在有待进一步研究的问题，包括：网格作业调度过程中资源抢占问题、数据部署约束下的网格作业调度问题以及如何将经济学机制的设计理论更深入地引用到网格经济模型之中等。更多还原

【Abstract】 With the development of computer hardware and network technologies it is possibble to integrate computing resources distributed in the wide area for distributed computing. From 1990 s, the grid technology has been concerned by various research institutions and business enterprises in the worldwide.By integrating heterogeneous, distributed and autonomous computing resources,the grid constructs a shared and cooperative computing environment.Job scheduling is an pivotal function of grid system. Because of the features of wide area in the grid environment, and the existence of inherent property of the grid resource nodes such as dynamic, autonomy and unreliability, job scheduling in grid environment is one of the key factors which constrain the development of the grid technology, as well as the emphasis and difficulty for the grid related researches. By introducing the basic principles of market economy in reality into the grid resources managment and job scheduling, the grid economy prompts the providers and users of the grid which are driven by the benefits to construct a large-scale grid enviroment. At present,grid economy has become a research focus of the grid and job scheduling.Based on the early related researches of the grid economy,this dissertation firstly analyzes and summarizes the related researches of the grid such as grid architecture, framework and job scheduling,and then some further researches, including the job type support, running risk of grid jobs, the pricing strategy of grid resources and the grid simulation are proposed. The mainly contributions and innovations in the dissertation are listed as below:1. A system-level job scheduling algorithm with the constraints of deadline and budget is proposed.The new algorithm improves the traditional application-level scheduling which only supports parameter sweep applications by supporting common parallel jobs. By considering the job initial priority and priority dynamic adjustment, it realized the support for the common parallel job scheduling in grid environment, and lays a foundation for the follow-up work.2. Proposing a risk control model which is based on the resources price in the process of job scheduling in grid environment. By introducing the utility of risk, it sloves the problem that how to characterize the users affordability of the risk in the grid which is an unreliable computing environment, while realizes the algorithm of the job scheduling in grid environment, which balanced the reliability of the grid node, the resource price and the users affordability for the risk.3. Proposing a multi-level resource pricing model of the single grid node. By jobs level divided and multi-level charged to the grid users, it solved the problem that the exsiting uniform pricing model can not distinguish different levels of the grid users, while realized to improve the profit of grid nodes ,as well as protecting the different level users qulity of service by settling the different resource scales.4. To determine the optimized resource scales of varied level users,the dissertation proposes an efficient leverage by using the theory of marginal expected benefit . Through setting up the demand distribution function of the resources by various level users, it solved the problem that how to determine the resources scales of the varied level users in the multi-level price, so that guaranteed the maximized expectation of the grid nodes profit.5. Extensions of GridSim, which is the most famous grid simulation tool in grid economy. GridSim can only simulated single CPU job, and is unable to characterize the serial portion of a common parallel application, as well as not supporting the third party data transfer and being lack of the resource distributed model under multi-CPU jobs. The extensions realizes the support for these most basic functions of grid simulation, which extends the application scale and flexibility of GridSim. The extended simulation tooll also supplies a suitable test platform for the related simulation tests of this dissertation.By using the extended simulation tool,some related tests are performed.The results show that by utlizing the related algorithms and models, grid jobs can be scheduled in more efficient manners in grid economy. The quality of service can be met as well as profit of grid nodes can be impoved.By using grid economy, the dissertation has a depth and meticulous researches in the areas of grid job system-level scheduling, job risk control,multi-level pricing model and the adjustment of realted resource scales. However, there are still some issues that are not be concerned for research, including the problem of resource preemption during the process of job scheduling in grid evironment, the problem of job scheduling in grid environment under the constrains of data deployment, as well as how to deeply introduce the basic theories of economics into grid economy model.更多还原