Tuesday, October 31
2:00pm,  SUBII Room 3014, Meeting Room 3

Low-Power Real-Time Computing

Abstract

With the advent of portable/embedded systems that rely on battery power, power/energy management has become a focus point of numerous research projects in the academia and industry. At the system level, existing techniques usually exploit various performance trade-offs to save power. In this talk, we will give an overview of our recent research on power management for real-time systems. A major component of our framework is the Dynamic Voltage Scaling (DVS) technique, recently incorporated to the processor lines of Intel and AMD. In essence, DVS consists in adjusting the CPU supply voltage and frequency (hence, the CPU speed) on-the-fly. While it is possible to obtain significant energy savings at the expense of increased response times with DVS, it is imperative that the timeliness guarantees be still provided for real-time applications.

We will first present an integrated real-time DVS framework, with static, dynamic and speculative solution components. All three components are formally shown to satisfy the timing constraints of tasks, while yielding remarkable energy savings. Then, we will focus on "energy-constrained" settings where the system has to remain functional and deliver an acceptable real-time performance during a given "mission time" with scarce and non-replenishable energy budget. We will present some fundamental intractability results, and show how efficient and practical solutions can be developed by adapting conventional real-time overload management techniques (such as reward-based and weakly-hard real-time computing models) to the energy-constrained settings. We will conclude with comments on the extension of the framework to multiprocessor systems.