RTHpower: Accurate fine-grained power models for predicting race-to-halt effect on ultra-low power embedded systems

Ultra-low power (ULP) embedded systems have become popular in the scientific community and industry, especially in media and wearable computing. In order to model ULP systems where energy per instruction can be as low as few pJ, more accurate fine-grained approaches are needed. However, there are no application-general, fine-grained and validated models yet that provide insights into how an application running on an ULP embedded system consumes energy and, particularly, whether the race-to-halt (RTH) strategy that are widely used in high-performance computing (HPC) systems is still applicable to ULP embedded systems. In this study, we propose new RTHpower models which provide insights into how an application consumes energy when running on an ULP embedded system. The models are trained and validated with data from 22 microbenchmarks. The experimental results show that RTH is not always applicable to ULP embedded systems, due to their low static power. RTHpower models support predicting when and when not to use RTH for a given application.