Power Profiling: From Measurements to Simulation Models

Abstract

Energy efficiency is currently one of the biggest challenges inmodern computer design. High power density limits further performancegrowth, and energy efficiency affects both the power bill for supercomputersand battery lifetime for embedded devices. A better understanding of energyefficiency during the design stage eases development of better architectures.In this thesis, we investigate energy consumption and architecturalproperties of an ARM Cortex-A9 processor. Further, this information is usedto create a tool for estimating its power consumption through simulation.Instruction level energy consumption is determined through measurements andexperiments on real hardware, which are further mapped to certainarchitectural events found in the gem5 simulator. The tool utilizes theseevents together with a simulator trace log and outputs a representation ofenergy consumption over time.This method can be applied during the development process at the simulatorlevel, while traditional methods typically involves hardware synthesis. Theresults show that this tool can estimate energy consumption with margin oferror of 5 % on general workloads, and is able to identify powerconsumption trends throughout a program.