Idle and Max Power Forecasting for Linear Server Power Consumption Model Using Random Forests
Ming Zhong, Dezhi Li, Tianheng Chen, Hongyin Chen, Ye Li, Xuejun Shang, Chenyang Zhao, Chao Liu · 2021 IEEE 5th Conference on Energy Internet and Energy System Integration (EI2) · 2021
The number of data centers is growing rapidly these years to meet the requirement of the development of internet technology. Data centers operate continuously throughout the year and consume a large amount of electricity every year. Consequently, they bring a great burden to the power departments and the environment. Optimization of workload scheduling in the data centers can reduce the power consumption of both internet technology system and cooling system. The power consumption model of servers is the basis of workload scheduling, and its accuracy and convenience are very essential to the implementation of scheduling. This paper focused on the linear power consumption model which proved to be simple and of high accuracy. Idle power and max power are two main parameters in the linear model and they can be measured through experiment tests. However, the experiment tests take much time and they can not be conducted on the servers already in use. Therefore, this paper aims to build a prediction model of both idle power and max power using readily accessible hardware parameters. A dataset crawled from the SPECpower website was firstly analyzed. And a random forests regression model was built based on it. The data set showed that idle power and max power had a downward and upward trend respectively. The idle power percentage, however, continued to decrease. The random forests regression model performed well. The mean average percentage error of idle power and max power was 16.7% and 10.3% respectively, which were much better than multiple linear regression model results and empirical equation results.