Research On Energy Efficiency Oriented Virtualized Resource Provisioning Method In Cloud

With the rapid development of cloud computing industry,energy consumption of large-scale data center is increasingly becoming a challenging issue.Industry and academia are trying to use a variety of techniques to improve the energy efficiency of data centers,and one of the most critical technology used in cloud computing is resource virtualization.Compared with the non-virtualized systems,the elasticity brought by virtualization can significantly improve the resource utilization,and achieve energy saving.However,competition for shared resources makes the performance of application within the virtual machines under threat.This directly affects the quality of cloud services,furthermore,the service provider's revenue and the end-user experience.How to achieve the best compromise between performance and power consumption on such a virtualized platform is one of the most concerned issues of industry.This dissertation focuses on virtualized resources provisioning method in cloud computing environments,and is intended to continually optimize the resource allocation of application with efficient dynamic resource provisioning policy,guarantee application performance,improve resource utilization in cloud computing data center,reduce data center energy consumption,and enhance system energy efficiency.The main contributions of this dissertation are as follows:1.Load forecasting method based on support vector machineVirtualization is the main technology to improve data center energy efficiency.In a cloud environment,many applications typically have highly variable load,resulting in complex and dynamic resource usage patterns.Only by real-time and dynamically adjusting virtual machine resource supply according to the actual needs of the application,can resource provisioning strategy minimize resource usage under the preconditions to meet QoS constraints.Because of the inevitable delay existing in the process of dynamic resource adjustment,a necessary precondition to achieve the above goal is to get the future resource requirements of the virtual machine in advance.This dissertation proposes a load forecasting method KSwSVR.It treats virtual machine load forecasting as a time series forecasting problems,and adopts the Statistical Learning Theory which is established specially for small sample statistical problem.The support vector machine is based on structure risk minimizing principle and has high generalization ability.Therefore,it can effectively adapt to the dynamic cloud computing environments.Inspired by the principle of locality,the paper improved the standard support vector regression algorithm data are weighted according to their importance and Kalman smoothing technology is integrated.Experimental results show that,KSwSVR can predict various types of resources,and outperform other commonly used classical algorithms in prediction accuracy,stability and cost.2.Automated resource scaling method based on load forecastingThe most important feature of cloud computing is elasticity.Application can acquire or release resource dynamically according to its own demand.From the perspective of the management system,the purpose is to make use of the dynamic resources scalability provided by virtualization,to make the resource allocation match the changing needs of application as much as possible.At the same time,the scale of the cloud computing system requires the scaling operation must be automated.It is needed to to reduce or even completely eliminate human intervention.Accurate load forecasting results can guide the development of the resource scaling scheme.However,because of the ubiquitous prediction error,directly taking the predicted value as allocation amount will lead to application performance instability.With the aim of guaranteeing application performance and maximizing resource utilization,this dissertation proposes a new automated resource scaling method G2 LC.It combines the experiment experience and QoS feedback,and adjusts the forecast result with global gain and local error compensation.While meeting the SLA constraints,it effectively reduces the unnecessary scaling operation caused by transient load,and completely avoids the concentrated emergence of SLA violation.Experimental results show that G2 LC can effectively guarantee any user-specified performance level by adjusting parameters.Compared with the fixed virtual machine resources method,G2 LC can significantly improve resource utilization,and this advantage become more apparent as user performance requirements increases.3.Resource provisioning based on QoS differentiationPerformance control and energy saving are the two main research themes of modern data center,but they are often conflicting.Cloud service providers use virtualization technology to integrate various applications onto fewer physical hosts,and convert idle host to low-power mode to reduce power consumption.In this case,the application performance is heavily dependent on efficient management of the virtual machine capacity.A variety of factors(such as application diversity,different usage patterns of various resources,shared underlying hardware,performance dependency and performance interference between applications)make the performance control difficult.It is critical to find a balance between power and performance.At the end of the dissertation,in order to improve the system energy efficiency,a new resource provisioning method is proposedCoST.It is based on QoS differentiation strategy,and takes advantage of the fact that different types of applications have different sensitivity to the performance and cost.Performance-sensitive applications pursue stable QoS,while performance-tolerant applications are more concerned about the total cost.CoST deploys different types of applications on the same host,and scale the resource vertically based on the load forecasting and QoS feedback.Experimental results show that CoST can effectively guarantee the QoS of performance-sensitive applications,support online modification of target performance metrics,and improve the overall processing speed of performance-tolerant application.The most important is that,CoST maintains the host always running in the most energy efficient state,improving overall system energy efficiency significantly.