Reliable Basic Block Energy Accounting.

Modeling the energy consumption of low-level code will enable (i) a better understanding of its relationship to execution time and (ii) compiler/runtime optimizations tailored for energy efficiency. But such models need reliable ground truth data to be trained on. We thus attack extracting machine-specific datasets for the energy consumption of basic blocks–a problem with surprisingly few solutions available. Given the impact of execution context on energy, we are interested in recording sequences of basic blocks coupled to corresponding energy measurements. Our design is lightweight and portable; no manual hardware/software instrumentation is required. Its main components are an energy estimation interface with sufficiently high refresh rate, access to an application’s complete execution trace, and LLVM pass-based instrumentation. We extract half a million basic block-energy mappings overall, and achieve a mean whole-program error of ∼ 3% on two different machines. This paper demonstrates that commodity resources suffice to perform a very crucial task on the road to energy-optimal computing.