Research On Cloud Resource Provisioning Strategy Based On Workload Prediction

Compared with the traditional application deployment,the application deployment of cloud computing has many advantages,such as scalability,cost-effective and high availability.Cloud computing provides pay-as-you-go model.While static resource provisioning is not flexible enough,dynamic resource provisioning is the first choice.But the traditional way of dynamic provisioning cannot avoid lagging problem.Only by predicting workload and resource demand in advance can the system provide enough resources to ensure application Quality of Service(QoS)in the specific scenarios of Web traffic burst.Meanwhile,it is necessary to minimize resource usage to reduce cost under the prerequisite of QoS guarantee for the application provider.Therefore,there is a need predicting the workload accurately.In order to ensure the availability of resources in Web traffic burst scenarios,this paper studies the research status of resource provisioning for Web cloud application.After the analysis of several typical workload prediction model,MASVR model is proposed,combined with the Moving Average(MA)model and Support Vector Regression(SVR)model.Combined with the Local Outlier Factor(LOF)algorithm to further improve MASVR model,workload prediction model named LOF-MASVR is proposed according to the specific characteristics of Web traffic burst scenarios.The 1998 World Cup workload dataset used in workload prediction experiment is to simulate the specific scenarios of traffic burst.Experimental results show that compared to MA,ARIMA and SVR prediction model,LOF-MASVR model can reduce Service Level Agreement(SLA)violation number in peak period nearly 70%.Moreover,prediction error of LOF-MASVR is significantly lower than that of ARIMA and SVR model,only slightly higher than that of MA model.While the resource allocation process takes 5?15 minutes,the prediction of LOF-MASVR model takes only a few seconds.It is obvious that the time efficiency of LOF-MASVR model is at an acceptable level.As existing prediction-based virtual resource provisioning mechanism using LOF-MASVR model is still unable to meet the availability requirements of application QoS.Combining with the LOF,and self-adaptive coefficient,this paper proposes a dynamic virtual resource provisioning strategy based on workload prediction and a corresponding cloud resource provisioning system architecture.The 1998 World Cup workload dataset used in resource provisioning experiment reveals the applicability of the system architecture and effectiveness of resource provisioning algorithm adapting to workload changes in the specific scenarios of traffic burst.Resource utilization and availability of proposed provisioning algorithm both meet the SLA requirements.Furthmore,proposed algorithm is significantly better than static provisioning strategy,and availability of proposed algorithm is better than that of existing dynamic provisioning strategy.