An efficient frequency scaling approach for energy-aware embedded real-time systems

The management of energy consumption in battery-operated embedded and pervasive systems is increasingly important in order to extend battery lifetime or to increase the number of applications that can use the system's resources. Dynamic voltage and frequency scaling (DVFS) has been introduced to trade off system performance with energy consumption. For real-time applications, systems supporting DVFS have to balance the achieved energy savings with the deadline constraints of applications. Previous work has used periodic evaluation of an application's progress (e.g., with periodic checkpoints inserted into application code at compile time) to decide if and how much to adjust the frequency or voltage. Our approach builds on this prior work and addresses the overheads associated with these solutions by replacing periodic checkpoints with iterative checkpoint computations based on predicted best-, average-, and worst-case execution times of real-time applications (e.g., obtained through compile-time analysis or profiling).