Prediction-based Resource Provision and Virtual Machines Placement in Cloud Data Center

Abstract

Abstract With the rapid development of virtualization, cloud servers provide the power and flexibility that single servers struggle to provide. However, low utilization of physical machines and high energy consumption are the main concerns for cloud service providers. In this paper, we propose a predict-and-place framework to decrease the number of active physical machines in cloud centers. We analyze the historical VM (virtual machine) request records to predict the VM demands in the next several days and then use an offline VM placement strategy to keep the number of active servers as less as possible while satisfying the SLA (Service-Level Agreement) requirements. In the prediction process, we enhanced the classical Holt’s linear prediction method on account of inevitable outliers to increase the prediction accuracy. We conduct experiments to compare the prediction accuracy of Perceptron, classical Holt-Winters, and our refined, robust Holt-Winters method. The experiment based on real data shows the simple perceptron has an inferior result. Our improved prediction method reduces the MAPE (Mean Absolute Percentage Error) by about 50% compared with the classical one, and the predict-and-place framework decreases the average number of active physical machines effectively in the cloud data center.