An analytical methodology to derive power models based on hardware and software metrics

摘要

The use of models to predict the power con-sumption of a system is an appealing alternative towattmeters since they avoid hardware costs and areeasy to deploy. In this paper, we present a systematicmethodology to build models with a reduced numberof features in order to estimate power consumption atnode level. We aim at building simple power modelsby performing a per-component analysis (CPU, mem-ory, network, I/O) through the execution of four stan-dard benchmarks. While they are executed, we collectinformation from all the available hardware countersand resource utilization metrics provided by the sys-tem. Based on correlations among the recorded metricsand their correlation with the instantaneous power, ourmethodology allowsi)to identify the significant met-rics; andii)to assign weights to the selected metrics inorder to derive reduced models. The reduction also aimsat extracting models that are based on a set of hardwarecounters and utilization metrics that can be obtainedsimultaneously and, thus, can be gathered and com-puted on-line. The utility of our procedure is validatedusing real-life applications on an Intel Sandy Bridge architecture.