Exploiting intra-task slack time of load operations for DVFS in hard real-time multi-core systems

Abstract

Power demand grows much faster than battery capacity in embedded systems. Dynamic voltage and frequency scaling (DVFS) has been shown to be extremely efficient to save energy due to the exponential dependence of power on voltage. However, voltage/frequency cannot be blindly scaled in hard real-time systems because DVFS techniques impact on the execution time, and so potentially on the worst-case execution time (WCET) of tasks. This paper presents a new DVFS technique for hard real-time systems that measures dynamically the intra-task slack existing between the actual execution time of a task and its WCET estimation, and exploits it to perform DVFS guaranteeing that the WCET is not affected. Concretely, our approach exploits the slack available due to contention in the use of shared resources in a multi-core system.