A high-level energy consumption model for heterogeneous data centers

Data centers consume anywhere between 1.7% and 2.2% of the United States' power. A handful of studies focused on ways of predicting power consumption of computing platforms based on performance events counters. Most of existing power-consumption models retrieve performance counters from hardware, which offer accurate measurement of energy dissipation. Although these models were verified on several machines with specific CPU chips, it is difficult to deploy these models into data centers equipped by heterogeneous computing platforms. While models based on resource utilization via OS monitoring tools can be used in heterogeneous data centers, most of these models were linear model. In this paper, we analyze the accuracy of linear models with the SPECpower benchmark results, which is a widely adopted benchmark to evaluate the power and performance characteristics of servers. There are 392 published results until October 2012; these servers represent most servers in heterogeneous data centers. We use R-squared, RMSE (Root Mean Square Error) and average error to validate the accuracy of the linear model. The results show that not all servers fit the linear model very well. 6.5% of R-squared values are less than 0.95, which means linear regression does not fit the data very well. 12.5% of RMSE values are greater than 20, which means there is still big difference between mod-eled and real power consumption. We extend the linear model to high degree polynomial models. We found the cubic polynomial model can get better results than the linear model. We also apply the linear model and the cubic model to estimate real-time energy consumption on two different servers. The results show that linear model can get accurate prediction value when server energy consumption swing in a small range. The cubic model can get better results for servers with small and wide range. The power requirements of today data centers range from 75 W/ft 2 to 150–200 W/ft 2 and will increase to 200–300 W/ft 2 in the nearest future. Energy cost becomes a major part of data center operational cost. To reduce the operational cost in large-scale data centers, researchers developed a wide range of energy-saving and thermal management techniques (e.g., workload consolidation, live migration, CPU throttling solutions). Workload consolidation is one of the most effective ways of conserving power by turning off spare servers. In many cases, the workload consolidation technique is incorporated with virtual machines, which are migrated from many physical 1569-190X/$-see front matter …