【作者】 车喜龙；

【导师】 胡亮；

【作者基本信息】 吉林大学 ， 计算机系统结构， 2009， 博士

【摘要】 资源的分配与任务的调度,是网格计算的核心功能,而资源性能无疑是最重要的影响因素。网格资源的性能信息主要通过两种方式来获取:网格资源监测与网格资源预报。本文工作主要围绕网格资源的监测与预报进行展开,讨论了系统架构的设计与实现,网格资源监测策略,网格资源建模预报,资源模型的优化等关键问题。主要包含以下几方面内容:(1)针对网格系统特性讨论了网格资源监测预报系统的设计原则,继而基于Web服务资源框架提出了系统架构,对其系统工作流程与系统信息管理进行定义。本系统开销较低,而且能够与网格计算系统实现无缝融合。(2)基于两类典型的计算智能方法——前馈人工神经网络与支持向量回归——对网格资源多步预报进行建模。实验证明这两类方法能够适用于网格资源建模,尤其是支持向量回归,在精度与效率上均表现出良好性能。(3)基于两类典型的计算智能方法——遗传算法与粒子群算法——对网格资源预报模型进行优化。实验证明,计算智能的优化方法性能优于改进后的传统搜索策略,尤其是本文提出的并行混合粒子群算法,在精度与效率上均表现出良好性能。并行混合粒子群优化Nu-支持向量回归模型能够满足在线优化与预报的要求,适用于网格资源监测预报系统。更多还原

【Abstract】 Grid computing breaks the limitations that exist in traditional shared environment, and becomes a leading trend in distributed computing system. Grid computing aggregates distributed and heterogeneous resources across Internet, regardless of differences between resources such as hardware structure,operating system, organization and geographical location. Such resources including computing resources, storage resources, data resources and other resources are combined dynamically to form high performance computing ability for solving problems in large-scale applications. From the point of view in system architecture, resource allocation and job scheduling is the core function of Grid computing, while there is no doubt that resource performance is an influencing factor of great importance.There are mainly two methods for acquiring performance information of Grid resources: Grid resource monitoring and Grid resource prediction. Grid resource monitoring cares about the running state, distribution, load and malfunction of resources in Grid system, by means of monitoring mechanism; Grid resource prediction focuses on the variation trend and running track of resources in Grid system, by means of modeling and analyzing historical monitoring data. Historical information generated by monitoring and future variation generated by prediction are combined together: to feed Grid system for analyzing performance, eliminating bottleneck, diagnosing fault, maintaining dynamic load balancing; and to help Grid user minimizing cost on time, space and money while acquiring task results.This research focuses on Grid resource monitoring and prediction, with main efforts dedicated on key issues including design and implementation of system architecture, monitoring strategies and modeling methodologies of Grid resources, optimization for resource model, etc.Grid resource monitoring and prediction derives from the seamless fusion of Grid technologies, resource monitoring strategies and resource prediction methodologies. Our research starts from Grid technologies. Characteristics and architecture of Grid system are firstly analyzed, then basic conceptions and relationships among them are illustrated, including Web service/Grid service, OGSA/OGSI (Open Grid Service Architecture/Infrastructure), WSRF (Web Service Resource Framework), etc. Furthermore, several typical resource monitoring and prediction systems are studied and introduced, then the monitoring strategies and prediction methodologies are summarized with comparison on their advantages and disadvantages. Such analysis and discussions are foundation of our research.Design principles for Grid monitoring and prediction system are given according to characteristics of Grid system, then the system architecture is proposed based on the WSRF. In order to reduce the influence on Grid system which is brought by of monitoring and prediction, overall architecture falls into two subsystems: resource monitoring subsystem inside computing environment and resource prediction subsystem outside computing environment. A series of supporting services are set up by way of Grid service technology. These services are deployed on Grid nodes and maintained by Globus service container. All functions of the monitoring and prediction system are accomplished through dynamical collaboration of the supporting services.These supporting services include monitoring service, prediction service, optimizing service, and information service. Monitoring service runs on computing node, it manages local resource sensors and generates resource monitoring data. Prediction service runs on prediction node, it manages resource prediction model and generates resource prediction data. Optimizing service runs on optimizing node, it evaluates efficiency and accuracy of candidate prediction model, and returns evaluation results to prediction service. Information service runs on resource database node, it is responsible for storage, query and publication of monitoring and prediction information. The workflow and information management of system are illustrated in details, while the performance is evaluated based on structure analysis and overhead tests.ANN (Artificial Neural Network) is born with parallelization, nonlinearity, robustness, and evolving capability. It breaks the limitation of traditional modeling methods, and becomes an important research category of computational intelligence. Feed forward neural networks are employed in modeling Grid resource prediction of multi-step-ahead. BPNN (Back Propagation Neural Network) is good at nonlinear mapping, while RBFNN (Radial Basis Function Neural Network) is good at clustering. These two neural networks are chosen for their wide applicability in practical problems. Furthermore, GHNN (General Hybrid Neural Network) is proposed which hybridizes RBFNN and BPNN together. Resource models based on three methods are compared during experiments. Results on efficiency and accuracy indicate that GHNN achieves lower error than BPNN and RBFNN in the cost of little increase on training time, and it is suitable for modeling Grid resource prediction of not only one-step-ahead but also multi-step-ahead.As a promising solution to nonlinear regression problems, SVM (Support Vector Machine) has recently been winning popularity due to its remarkable characteristics such as good generalization performance, the absence of local minima and sparse representation of the solution, thus is expected to achieve better performance. We discussed the modeling issues of SVR (Support Vector Regression), including modeling steps, choice of kernel functions, solution to QP (Quadratic Programming) problems, etc. Epsilon-SVR and Nu-SVR are realized as two typical algorithms in SVM. Their performance are tested in experiments and compared with GHNN. There is a performance similarity between two SVR models with default parameters, and they achieve better performance than GHNN according to comparative results on efficiency and accuracy. SVR is validated as a proper method in modeling Grid resource prediction of both one-step-ahead and multi-step-ahead, thus is chosen in building prediction system.After choosing modeling method, we further discuss the optimization issues of Grid resource prediction model of multi-step-ahead. Traditional search strategy plays the roles of optimizing method in many practical applications. However, this method has high complexity on time and space, thus is unfeasible for online system. TSS (Two Stage Search) strategy is proposed based on improvements to original strategy using granularity control mechanism. The optimizing performance of TSS strategy is validated through experimental results. It can’t break the limitation of enumeration in despite of accuracy enhancement to prediction model. The optimizing time is over 60 seconds, which can’t meet the efficiency demand for online system. Therefore, better optimizing method is also required.GA (Genetic Algorithm) and PSO (Particle Swarm Optimization) are two typical stochastic optimization methods in the category of evolutionary computation. These two methods are employed to optimize SVR prediction model, for the expectation of achieving higher performance. Parameter selection is taken under consideration and TSS is taken under comparison. Experimental results indicate that the efficiency and accuracy of both GA and PSO are better than TSS. PSO achieves lower error and less optimizing time than GA, and has less prematurity phenomenons.Due to the superior performance of PSO, a PH-PSO (Parallel Hybrid Particle Swarm Optimization) algorithm is proposed which hybridizes discrete PSO and continuous PSO together, for the purpose of combinational optimization to Nu-SVR prediction model, including feature selection as well as hyper-parameter selection. Comparative results indicate that the PH-PSO algorithm has high efficiency and good convergence. Furthermore, the Nu-SVR model optimized by PH-PSO can achieve high accuracy in Grid resource prediction of multi-step-ahead. Besides, the optimizing time has been remarkably reduced to less than 3 seconds. The combinational model of PH-PSO and Nu-SVR meets the accuracy and efficiency demand of online system, thus is suitable for building Grid resource monitoring and prediction system.The results of the dissertation will contribute to the building and advancing of Grid infrastructure. In the next steps, our research will go further in the following applications: monitoring and prediction of Grid tasks, classification and evaluation of Grid resources, classification and evaluation of Grid tasks, etc. It is believed that as important tools for modeling and optimizing, computational intelligence will play a more important role by the virtue of its potential in the field of Grid computing.更多还原