Energy-Driven Statistical Sampling: Detecting Software Hotspots

Energy is a critical resource in many computing systems, motivating the need for energy-efficient software design. This work proposes a new approach, energy-driven statistical sampling, to help software developers reason about the energy impact of software design decisions. We describe a prototype implementation of this approach built on the Itsy pocket computing platform. Our experimental results of 14 benchmark programs show that when multiple power states are exercised, energy-driven statistical sampling provides greater accuracy than existing time-driven statistical sampling approaches. Furthermore, if instruction-level energy attribution is desired, energy-driven statistical sampling may provide better resolution. On simple handheld systems, however, many applications may exercise only a single power state other than idle mode. In this case, time profiling may sufficiently approximate energy profiling for the purpose of assisting programmers, without requiring any hardware support.