Abstract:
Energy efficiency and reliability management are two important aspects of real-time embedded systems, in addition to strict guarantees for timing constraints. The objective of this research is to optimize performance metrics such as energy consumption and quality-of-control, by means of appropriately scheduling computational tasks and applying system- level energy management techniques. Reliability is achieved via run-time fault tolerance, which must be based on additional hardware and/or software components. These additional components are often redundant for normal operation and they can cause significant increase in overall energy consumption. Therefore, simultaneously managing both energy and reliability, two opposing factors, has been an intriguing research topic for real-time embedded systems. In this dissertation, we first address the problem of minimizing energy consumption for the emerging heterogeneous multi-core systems, which have a wide range of power-performance characteristics. Reliability is achieved via scheduling two copies (primary and backup) of each task on different processing cores. Each processing core is equipped with Dynamic Voltage and Frequency Scaling and Dynamic Power Management features in order to reduce the energy consumption. We address the problem for independent real-time tasks on a heterogeneous dual-core processor, and we propose several algorithms which take advantage of the heterogeneity to reduce energy consumption while ensuring hard deadlines with reliability guarantees. We also address the problem for tasks with precedence constraints and develop efficient solutions. Then, we tackle the problem for the general real-time periodic task model with arbitrary periods for the same heterogeneous platform. We propose techniques for task partitioning, priority assignment, and runtime scheduling of periodic tasks to achieve reliability and energy efficiency. Evaluation of the proposed schemes was conducted by extensive simulation experiments which show their effectiveness within a wide range of system parameters. Finally, we tackle the problem of partitioning a set of real-time tasks on a homogeneous multiprocessor control system, and at the same time, assigning periods to those tasks (within an allowable range) in order to maximize the quality-of-control for the system. We model the quality-of-control as a concave function of task invocation rates (periods). We propose a family of heuristics that are based on effectively converting the multiprocessor problem to a single but faster uniprocessor system. We conduct extensive simulation experiments that demonstrate the superior performance of our proposed algorithms.
