Towards energy-efficient linear algebra with an ATLAS library tuned for energy consumption

Abstract

Autotuning is an established method for adapting the execution of an application to the underlying hardware for minimising the execution time. This article investigates whether autotuning is also suitable for minimising the energy consumption of an application. The investigation is done with the linear algebra library ATLAS. Adaptations for the ATLAS package which enable energy autotuning are proposed. Different tuning parameters are investigated for whether they show a different behaviour when ATLAS is tuned for energy consumption instead for execution time. The results suggest that some tuning parameters have to be set differently when ATLAS is supposed to work with a minimum energy consumption than with a minimum execution time. The results further indicate that tuning the complete ATLAS package for energy consumption leads to a more energy-efficient execution than tuning it for execution time.