CSR: Small: ENACT: Environment-Aware Management of Mobile Systems

University Of California-San Diego, La Jolla CA

Investigators

Abstract

Mobile systems are pervasive and have been adapted for diverse computing needs throughout the world. Smart mobile systems, with advanced computing and connectivity, promise superlative experience, yet battery constraints limit their use. Device temperatures are also a major concern as high temperatures reduce hardware reliability but also because human skin can tolerate only moderately warm phones. Mobile system usage is highly influenced by user context, hence can provide a crucial input for proactive, instead of reactive, device power, thermal and reliability management. Even with a small amount of contextual information such as user location and motion, battery lifetime can be improved by as much as a factor of five. Yet, despite the fact that there are already many sensors available in mobile systems, there have been no simple and flexible ways to incorporate context into mobile system resource management. To enable mobile systems to become more context-aware, this project is designing ENACT, an ENvironment Aware ConTrol framework that leverages overarching sensor data for system-wide context-aware management. ENACT is a lightweight framework that enables a mobile system to tap into a vast array of sensors, and leverage comprehensive context about hardware, software, and user, to control system-wide actuators. The framework has two main components. A context recognition service leverages well-established statistical techniques to robustly derive semantic information from raw sensor data. Then, by observing system behavior with respect to context and actuators, it computes stochastic models. A context-aware control service, based on stochastic model predictive control, smartly sets the system's actuators to minimize the energy consumption while meeting reliability, temperature, and performance constraints. To efficiently meet these requirements, the control framework is supported by a hierarchical structure working at different time scales. ENACT can efficiently transforming a large set of raw sensor data into usable contextual information accessible system-wide, and leverage context to configure a set of actuators to improve the way mobile system resources are managed while providing excellent user experience. This allows for accurately aligning system goals to performance expectations. The project includes the implementation of novel context-aware techniques for energy, thermal and reliability management, which can smartly adapt to dynamic performance requirements without sacrificing energy margins. Students involved in this project gain valuable training in disciplines across machine learning, hierarchical control, optimization and resource management. The ENACT framework can be leveraged in academia and industry alike to implement context-awareness in systems, and to design resource management policies that can leverage context.