Optimal state estimation for improved power measurements and model verification: Theory

Abstract

To improve energy efficiency in computer systems and data centers, accurate models of the power consumption are needed for analysis and advanced control algorithms. Developing models requires deep understanding not only of the components themselves but also their interaction. Moreover, verifying models requires accurate measurements, which in itself requires some understanding of the system. Optimal state estimation is a well established field, which comprises mathematical methods often used for sensor fusion and to handle measurement inaccuracies. Optimal state estimators combine various measurements and the physical model of the system to acquire more accurate information. Optimal state estimation can also be used to test and verify different kinds of models and to identify system parameters. These algorithms also fit well to computer environments, making them a viable candidate for use in various on-line modeling, analysis and control techniques. This paper investigates the use of optimal state estimation to verify and improve system models. A simplistic model is first derived for a typical data center powering structure including cooling system. Multiple model and parameter identifying estimators are then proposed for validating the model and estimating model parameters. Theory presented in this paper has been formulated to enable accurate measurements as well as component and system-level model analysis in an upcoming data center test facility, currently under construction.