Research On Component Service Response Time Prediction In Cloud Based On FA-ELM And CRF

Cloud computing is a computing mode that faces the service-oriented characteristics and provides different resources to users in the form of services.What the cloud service users most care about is whether the services performance meet the signed SLA or not.For the cloud service providers,they consider achieving resource flexibility on the basis of meeting the performance requirements of users.Response time is one of the most important parameters in the SLA signed by service providers and users,and it is one of the performance indexs that users most concerned about.Cloud services can be composed of multiple component services that run on different virtual machines,the response time of the component services determine the response time of cloud services,so we need to study on the prediction method of component services response time based on the states of the virtual machine resources,providing strong supports for service providers to elastically adjust resources timely so as to satisfy the signed SLA with users.The research of component service response time prediction in cloud environment has become one of the hottest spots in Cloud Computing.Aiming at response time prediction of component services based on the resource states in virtual machines,this thesis puts forward two response time prediction methods from the view of both the single value prediction of response time and the series values prediction of response time.Single value prediction of response time is predicting the response time based on the current resources states.For the sequence prediction of response time,we not only consider the states of the current system resources,but also the influence of system resource states before and after a period of time on response time.For the single value prediction of response time,this thesis puts forward the fireworks optimized algorithm for extreme learning machine(FA-ELM)method,which builds a mapping between virtual machine resources states and component service response time based on ELM,and then predicts response time based on the mapping.Against the disadvantages of ELM,we proposed Fireworks Algorithm to select the input layer weight matrix and the implicit bias of ELM optimally.Compared to ELM,FA-ELM has greatly improved the prediction ability and stability.For the sequences value prediction of response time,we not only consider the current state of system resourcrs,but also consider the influence of dynamic changes of resources states on response time in the virtual machines,establishing a mapping relationship between response time series and sequence changes of resources states in virtual machines based on CRF,predicting the change trend of response time of component services based on the mapping relationship over a period of time in the future.Experients show that the method is effective.Finally,taking the initial allocation of resources in cloud environment for example to illustrate the application of the proposed methods in this paper.