Adaptive Workload Forecasting in Cloud Data Centers

Abstract

Forecasting on different levels of the management system of a cloud data center has received increased attention due to its significant impact on the cloud services quality. Making accurate forecasts, however, is challenging due to the non-stationary workload and intrinsic complexity of the management system of a cloud data center. It is possible to prevent excessive resource allocation and service level agreement violations through workload forecasting for virtual machines and containers. In this paper, the authors propose the adaptive forecasting model and corresponding adaptive forecasting methods to apply in the management system of a cloud data center for workload forecasting, ensuring compliance with the service level agreement and power consumption decrease. The authors consider six alternative forecasting methods and 77 training data windows on each management step to determine the best combination of methods and the training set size that generates a most accurate forecast, thereby adapting to the current state of the physical or virtual server in a cloud data center. Through the comprehensive analysis, the authors also evaluate the proposed adaptive forecasting methods using real-world workload traces Bitbrains and demonstrate that combined forecasting methods outperform the individual forecasting methods significantly in terms of forecasting accuracy measured by Mean Absolute Percentage Error.