Joint Reliability and Energy Management for Real-Time Embedded Systems

Abstract

The Dynamic Voltage Scaling (DVS) technique is the basis of numerous state-of-the-art energy management schemes proposed for real-time embedded systems. However, recent research has illustrated the alarmingly negative impact of DVS on system reliability, in terms of increased vulnerability to transient faults, leading to soft errors. The main theme of this dissertation is to investigate several open problems and trade-off opportunities in energy, reliability, and timeliness dimensions. This dissertation, first, investigates the problem of maximizing the overall reliability of real-time embedded applications while meeting the deadlines and a given energy budget constraint. Optimal static solutions and effective dynamic (online) solutions are developed. Second, the dissertation proposes a new approach, called the shared recovery (SHR) technique, to minimize the system-level energy consumption while still mitigating the reliability loss induced by DVS. The main idea of the SHR technique is to avoid the off-line allocation of separate recovery tasks to the scaled tasks by assigning a global/shared recovery block that can be used by any task at run-time. Specifically, an array of reliability-aware energy management algorithms are presented for both independent and dependent tasks. Third, the dissertation presents the foundations of a general reliability-oriented energy management framework, where the objective is to achieve any reliability level with minimum energy consumption and timeliness guarantees. For periodic real-time tasks, the framework is extended to address multiple reliability objectives that can be set by the designer and vary from task to task. Finally, this dissertation considers the problem of minimizing the expected energy consumption for a real-time embedded application. This part of the research integrates optimally DVS and Dynamic Power Management (DPM) techniques that can put some system components to sleep states when they are not in use for the case where the workload is known only probabilistically.