A classification-based demand trend prediction model in cloud computing

Abstract

Cloud computing allows dynamic scaling of resources to users as needed. With the increasing demand for cloud service, a challenging problem is how to minimize cloud resource provisioning costs while meeting the user’s needs. This issue has been studied via predicting the resource demand in advance. Existing predicting approaches formulate cloud resource provisioning as a regression problem, and aim to achieve the minimal prediction error. However, the resource demand is often time-variant and highly unstable, the regression-based techniques can not achieve a good performance when the demand changes sharply. To cope with this problem, this paper proposes a framework of predicting the sharply changed demand of cloud resource to reduce the VM provisioning cost. In this framework, we first formulate the cloud resource demands prediction as a classification problem and then propose a robust prediction approach by combining Piecewise Linear Representation and Weighted Support Vector Machine techniques. Our proposed method can capture the sharply changed points in the highly unstable resource demand time series and improves the prediction performance while reducing the provisioning costs. Experimental evaluation on the IBM Smart Cloud Enterprise (SCE) trace data demonstrates the effectiveness of our proposed framework.