云计算环境下资源需求预测与优化配置方法研究

发布时间：2018-06-18 01:44

本文选题：云计算 + 资源管理　；参考：《合肥工业大学》2014年博士论文

【摘要】：云计算是一种基于互联网的新型信息资源服务系统，可以为用户提供包括基础设施、平台和应用在内的可定制弹性虚拟化资源服务。在技术进步、需求引领和服务模式创新等因素的共同驱动下，云计算得到了工业界和学术界的普遍认可，已经在现实生活中形成涵盖移动互联网、物联网等在内的新型创意产业，并以其低成本和无处不在的应用得到迅速发展，将从根本上改变人们生活的方方面面。为了满足这些多元化、海量的应用资源需求，云计算必须拥有庞大的资源集群，这些资源在地理上是分布的，类型上是异构的，并且在各自的管理域中又具有不同的资源管理策略和资源使用计价准则。资源管理是云计算的核心问题之一，其目的是利用虚拟化技术屏蔽底层资源的异构性和复杂性，使得海量分布式资源形成一个统一的巨型资源池，并在此基础上，合理运用相关资源管理方法和技术，确保资源的合理、高效的分配和使用。因此，如何实现对云计算资源的有效管理成为一个富有挑战性的研究课题。本文从云计算基础设施运营商和服务提供商的角度出发，将研究内容集中于云计算资源的优化管理方向，主要包括资源的描述、组织、发现、匹配、配置和监控等内容，着重研究了云计算资源负荷的短期动态预测方法，基于短期预测的云资源优化配置方法，以及云资源需求中长期组合预测方法。目的是通过对以上几方面内容的研究，使得云计算资源能够得到有效地组织和合理的配置，在保证资源服务质量的同时，降低数据中心能源消耗和运营成本，提升云计算基础设施运营商和服务提供商的利润，实现绿色计算，为云计算的健康、持续发展提供理论参考。基于以上论述，本文云计算环境下资源需求预测与优化配置方法研究的主要内容有：云计算环境下资源管理问题综合研究；基于特征提取与分类的云资源负荷短期动态预测方法研究；基于云计算负荷短期动态预测的资源优化配置方法研究，以及针对云计算基础设施运营商和服务提供商中长期资源总量规划需求的云资源需求中长期组合预测方法研究。本文的具体研究内容和创新性工作主要有以下几个方面：首先，在总结了以往云计算资源描述格式和语言、发现架构和技术，以及动态组织、优化分配和即时监控等方面研究成果的基础上，进而阐述了云环境下资源管理所面临和需要解决的新问题，并以此构建了云环境下资源管理框架，给出了该框架在制造业背景下的应用思路。其次，分析了云计算资源需求负荷相对于先前的网格计算、分布式计算及其它高性能计算所表现出不同特点的基础上，讨论了短期负荷预测对于云计算实现资源实时控制、保持整个系统稳定运行、降低数据中心能耗和保障云服务的QoS所起的重要作用，构建了基于资源负荷序列特征提取、分类和预测的多步骤预测方法。该方法运用定长重叠移动滑窗技术从云计算资源负荷序列中提取子序列，再分别利用基于核模糊C聚类的监督式聚类算法和基于隐形马儿科夫链的非监督式聚类算法对所提取的子序列进行特征分类，在此基础上，再使用基于遗传算法优化的Elman神经网络对云计算短期动态资源负荷进行预测，以此获得优良的预测效果。接着，基于云计算短期负荷预测的结果，本文构建基于负荷预测的云计算资源优化配置框架，提出了一种基于资源监控和负荷预测的资源配置自适应弹性控制系统，实施主动控制与被动反应相结合的混合弹性控制的资源配置策略以实现云计算资源的有效利用；进一步地，，鉴于目前云计算服务提供商所采用的单虚拟机服务单用户的资源管理模式所带来的低资源利用率问题，本文构建了一个具有五层结构的新型公有云架构，在该架构的基础上，提出了基于单虚拟机服务多用户的虚拟化资源自适应配置模式，该模式能针对不同用户提出的应用资源请求自动搜寻最优虚拟化资源，并在不影响服务质量的基础上，将不同的应用运行在同一台虚拟机上，使得云计算提供商能在保证服务质量的同时，提高云计算资源的利用效率，降低能耗。最后，本文根据实际云计算资源管理中对资源负荷中长期预测的需求，针对云计算中长期负荷所表现出的兼具动态性和周期性这一特点，构建了基于广义模糊软集理论的云计算资源负荷组合预测模型，提出了新的基于夹角余弦的广义模糊软集相似性度量方法，将相似性度量结果与预测精度相结合来获得各单项预测模型的权重，并针对云计算环境中资源负荷所表现出的短期动态性和长期周期性特征，选用自适应神经模糊推理系统ANFIS（Adaptive Neuro FuzzyInference System）和季节性ARIMA模型SARIMA作为单项预测模型来分别处理其动态性和周期性特征，以此构建基于广义模糊软集理论的云计算资源负荷组合预测模型GFSS-ANFIS/SARIMA。
[Abstract]:Cloud computing is a new information resource service system based on the Internet, which can provide customizable flexible virtual resource services, including infrastructure, platform and application. Under the common drive of technological progress, demand guidance and service mode innovation, cloud computing has been widely recognized in industry and academia. In real life, the new creative industries, including the mobile Internet, the Internet of things and so on, have been developed rapidly with its low cost and ubiquitous applications. It will fundamentally change all aspects of people's life. In order to meet these diversities, the massive resources need to be used, cloud computing must have a huge collection of resources. The resource management is one of the core problems of cloud computing. The purpose of the resource management is to shield the heterogeneity and complexity of the bottom resources by virtualization technology, so that the mass distribution is distributed. Type resources form a unified huge pool of resources, and on this basis, the rational use of related resources management methods and technologies to ensure the rational and efficient allocation and use of resources. Therefore, how to realize the effective management of cloud computing resources has become a challenging research topic. The research content concentrates on the optimization management direction of cloud computing resources, mainly including the description, organization, discovery, matching, configuration and monitoring of resources, focusing on the short-term dynamic forecasting method of cloud computing resource load, the method of cloud resource optimization based on short-term prediction, and the requirement of cloud resources. The aim of the medium and long term combination forecasting method is to make the cloud computing resources effectively organized and reasonably configured through the study of the above aspects, and to reduce the energy consumption and operation cost of the data center while guaranteeing the quality of the resources service, and improve the profits of the cloud computing infrastructure operators and service providers. Green computing provides a theoretical reference for the healthy and sustainable development of cloud computing.
Based on the above discussion, the main contents of the research on resource demand forecasting and optimal configuration under cloud computing environment are: comprehensive research on resource management in cloud computing environment; research on short-term dynamic prediction method of cloud resource load based on feature extraction and classification; resource optimization based on short-term dynamic forecasting of cloud computing load Method research and long-term combined forecasting method for cloud resource demand in cloud computing infrastructure operators and service providers.
The specific research contents and innovative work in this paper are as follows:
First, on the basis of summarizing the previous description format and language of cloud computing resources, and finding the results of architecture and technology, dynamic organization, optimal allocation and real-time monitoring, this paper expounds the new problems facing and needs to be solved in resource management under the cloud environment, and builds a framework of resource management under the cloud environment. The framework of the framework in the context of manufacturing applications.
Secondly, based on the different characteristics of the previous grid computing, distributed computing and other high performance computing, the demand load of the cloud computing resources is discussed, and the short-term load forecasting for the real-time control of the resources for the cloud computing, keeping the whole system running steadily, reducing the energy consumption of the data center and ensuring the QoS of the cloud service is discussed. The multi step prediction method based on resource load sequence feature extraction, classification and prediction is constructed. The method uses fixed length overlapping mobile sliding window technology to extract subsequences from cloud computing resource load sequence, and then uses supervised clustering algorithm based on Kernel Fuzzy C clustering and non supervision based on stealth horse paediatrics chain. On the basis of this, the Elman neural network based on genetic algorithm is used to predict the short-term dynamic resource load of cloud computing, so as to obtain good prediction results.
Then, based on the result of cloud computing short-term load forecasting, this paper constructs the framework of cloud computing resource optimization based on load forecasting, proposes a resource allocation adaptive elastic control system based on resource monitoring and load forecasting, and implements the resource allocation strategy of hybrid elastic control with active control and passive reaction. In this paper, a new public cloud architecture with five layers of structure is constructed in view of the low resource utilization problem brought by the single virtual machine service single user resource management model adopted by cloud computing service providers. On the basis of this architecture, the paper proposes a single virtual machine based on the single virtual machine. This model can automatically search for the optimal virtual resources for the application resources proposed by different users, and on the basis of the quality of service, the different applications run on the same virtual machine, so that the cloud computing provider can improve the quality of service while improving the quality of service. The efficiency of the use of cloud computing resources to reduce energy consumption.
Finally, according to the requirement of long term prediction in the management of resource load in the management of cloud computing, this paper constructs a cloud computing resource load combination forecasting model based on the generalized fuzzy soft set theory, and proposes a new broad sense cosine based on the characteristics of the long and long term load in the cloud computing. The fuzzy soft set similarity measure method combines the similarity measurement results with the prediction accuracy to obtain the weight of each single prediction model. The adaptive neural fuzzy inference system ANFIS (Adaptive Neuro FuzzyInference System) is selected for the short-term dynamic and long-term periodic characteristics of the resource load in the cloud computing environment. The seasonal ARIMA model SARIMA is used as a single prediction model to deal with its dynamic and periodic characteristics respectively, so as to construct the cloud computing resource load forecasting model GFSS-ANFIS/SARIMA. based on the generalized fuzzy soft set theory.
【学位授予单位】：合肥工业大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：TP393.07

【参考文献】