The explosive growth in the size of data centers, coupled with the wide-spread use of virtualization technology has brought power and energy consumption as major concerns for data center administrators. Provisioning decisions must take into consideration not only target application performance but also the power demands and total energy consumption incurred by the hardware and software. Failure to do so will result in damaged equipment, power outages, and ine_cient operation. Since database servers comprise one of the most popular and important server applications deployed in such facilities, it becomes necessary to have accurate cost models that can predict the power and energy demands that each database workloads will impose in the system. In this paper we present an empirical methodology to estimate the power and energy cost of database operations. Our methodology uses multiple-linear regression and factorial experimental design to derive accurate cost models that depend only on readily available statistics such as selectivity factors, tuple size, numbers columns and relational cardinality. Moreover, our method does not need measurement of individual hardware components, but rather total power and energy consumption measured at a server. We have implemented our methodology, and ran experiments with several server configurations. Our experiments indicate that we can predict power and energy more accurately than alternative methods.
