Application of Machine Learning Algorithms for Predicting Energy Consumption of Servers

Abstract

Energy management in data centers is currently a major challenge and arouses considerable interest. Many data center operators are seeking solutions to reduce energy consumption. In this work, the problem of resource overutilization-defined as the excessive usage of critical server resources such as CPU, RAM and storage surpassing their optimal capacity-in data centers is addressed, with a particular focus on servers. Estimating the energy consumption of servers in data centers allows its managers to allocate the necessary resources to ensure adequate quality of service. The research involved generating workloads performance on various servers, each connected to a wattmeter for energy consumption measurement. Data on resource utilization rates and server energy consumption were stored and analyzed. Machine learning models were then used to forecast server energy consumption. Parametric, nonparametric, and ensemble methods were employed and validated using accuracy measurements, non-parametric tests, and model complexity to assess the quality of energy consumption prediction models. The results demonstrated that certain models could provide predictions with a low margin of error and minimal complexity like polynomial regression, while other models showed lower performance. A comparative analysis is conducted to evaluate the performance and limitations of each approach.