Abstract
Cloud computing infrastructure is designed to deploy and assess service-oriented applications, primarily via cloud datacenters. These datacenters are integral to energy utilization in cloud environments, with energy consumption closely tied to resource utilization. It is important to monitor and predict power consumption in these datacenters, especially for high-demand services. Container-based virtualization, particularly using Docker containers, has gained significant attention due to its lightweight nature. However, predicting energy usage at a fine-grained level for container-based applications is a challenging task. In this study, we employ three time series analysis algorithms—AR, ARIMA, and ETS—to predict the energy usage of Docker containers over the next hour. Utilizing collected time-series power consumption data, our study contributes to enhancing power predictions for Docker containers within cloud infrastructures. Our prediction results focus on four Docker containers, each running multiple applications as Docker subprocesses. Power data for individual applications was aggregated to determine total container power consumption. Comparing the performance of ARIMA, ETS, and AR algorithms in predicting Docker container instance power, we found varying outcomes across containers. Through assessing MAPE across different time series model window lengths, we identified superior performance among the models. Specifically, ETS consistently demonstrated the lowest MAPE values for containers like ’polinx-container’ and ’alpines-container’, indicating higher prediction accuracy compared to ARIMA and AR models. The ARIMA model outperformed the ETS and AR models for the ’progrium container’. These findings underscore the necessity of selecting appropriate time series models tailored to specific Docker container configurations and workload scenarios for precise energy consumption forecasts.