Online Dynamic Voltage Scaling in Embedded Real-Time Systems

Embedded systems are everywhere, from mobile phones to medical devices. They are becoming more complex, and have to fulfill competing requirements, in terms of performance (timing constraints), reliability and energy consumption (long battery life). Due to the competing constraints, their design is becoming increasingly difficult. In this thesis we consider hard real-time applications mapped on distributed heterogeneous architectures. The applications are modeled as a set of tasks, which are characterized by a worst-case execution time and a deadline. The processors in the heterogeneous architecture have have multiple operating modes, consisting of a given frequency and voltage. We assume that we know the reliability of the architecture components. Embedded systems are often designed using static approaches, where the implementation is derived offline. However, many applications require more flexible solutions. Hence, researchers are advocating adaptive approaches, which can change the system configuration in response to changes in the requirements and the environment. The objective of this thesis is to design and implement online adaptive scheduling algorithms, which are able to successfully address competing design objectives in terms of performance, energy consumption and reliability. We have adapted performance, energy and reliability model from the literature. The trade-off between performance and energy consumption has been addressed using dynamic voltage scaling (DVFS), i.e., reducing the dynamic power consumption by scaling down operational frequency and circuit supply voltage. However, lowering the voltage to reduce the energy consumption will impact negatively the reliability, and we plan to investigate also this energy/reliability trade-off. We have designed and implemented a simulation framework. Two main algorithms have been designed, implemented and compared: Low-Power Priority-Based Rate Monotonic (LPP) and Cycle-Conserving Rate Monotonic (CCRM). The simulation framework has been implemented using Java and multiple third party libraries. The framework is extensible, and we have shown that it can used to successfully evaluate the quality of online scheduling algorithms.