Toward an evaluation infrastructure for power and energy optimizations

Execution-driven simulators are often used for power/energy and performance evaluation. Simulators can provide semantic details but they provide insufficient speed and accuracy for compiler and OS research. Physical measurement is fast and objective but lacks a semantic connection between the measurement result and the evaluated program. The objective of our research is to bring together the advantages of simulation and physical measurement to build an infrastructure for power and energy optimization. Power and energy behavior is obtained through physical measurement. Simulation is used for observing the connection between power and energy behavior and the evaluated program. Our preliminary results demonstrate the ability of this infrastructure to capture detailed power behavior of any region of a program. To simplify the power/energy evaluation of programs with long execution times and overcome the limitation of physical devices, we propose using the SimPoints methodology developed by researchers at UC San Diego to find representative slices of a program. Through simulation, we validate the feasibility of the SimPoint idea in simplifying power/energy evaluation. We expect that this infrastructure will help researchers in OS/compiler power/energy optimization to evaluate their optimizations more efficiently and observe more optimization opportunities. 1