Evaluating techniques for method-exact energy measurements: towards a framework for platform-independent code-level energy measurements

Reducing the energy consumption of executing programs decreases the operational cost and the environmental impact. For some well-known tasks, domain-specific guidelines for energy-efficient code exist. However, we want to enable profiling the energy consumption of code and selectively optimize hot spots, previously unknown. On our way to providing method-exact information about energy consumption in different hardware components, we currently focus on measuring the CPU's energy-consumption due to running a program itself. We identify three main challenges: measuring energy consumption at a fine granularity, constructing repeatable measurements, and mapping measurements to code. By actually building and using a platform-independent measurement setup, we show that it is feasible to measure the energy consumption at the granularity of method executions. The measurement time series must be aligned with execution events such as relevant method executions. We demonstrate how to evaluate alignment techniques by the examples of linear scaling and Dynamic Time Warping for mapping the measurements to code. While both are not yet sufficient to reliably map measurements to fine-grained code locations, our framework demonstrates the general feasibility of our goal and enables further research in this direction.