Source-Level Estimation of Energy Consumption and Execution Time of Embedded Software

Abstract

Energy optimization of embedded software is of primary importance. Nevertheless, there is lack of accurate and usable methodologies and tools to estimate software performance (execution time, energy) and to allow a significant exploration of design alternatives. Current approaches use either instruction-level simulation (accurate but slow), or static-time source characterization (flexible but data-independent). This paper proposes a hybrid approach taking advantage of the strengths of both the above approaches. We present a fully automatic method for estimating the execution time and power consumption of a C program - run on a given architecture on given input data - based on statistically-accurate models for the architecture and for the compiler. Validation results against an ARM energy-enabled instruction-level simulator show an average absolute relative errors of 8.5%.