A Self-Organizing Wireless Sensor Network and Distributed Computing Engine for Commodity and Future Palmtop Computers

摘要

The embedded class processors found in commodity palmtop computers continue to become increasingly capable while retaining an energy-efficient footprint. Palmtop computers themselves, including smartphones and tablets, provide a small form factor system integrating wireless communication and non-volatile storage with these energy-efficient processors. Also, various wireless connectivity functions on mobile devices provide new opportunities in designing more flexible, smarter wireless sensor networks (WSNs), and utilizing the computation power in a way we could never imagine before. In this dissertation, I present a WSN concept for current and future generation tablet devices. My contributions include developments at the system level, architecture level, and collaborative design between different layers of the system. At the system level, I developed Ocelot, a grid-like computing environment for palmtop computers in place of traditional workstation or server class machines to compute highly parallel light to medium-weight tasks in an energy efficient manner. Additionally, I developed Lynx, a self-organizing wireless sensor network, which is a further step taken in exploiting the potential of palmtop computers. At the architecture level, to increase the storage capacity of future palmtop computers, I explore the use of a new storage class magnetic memory, Racetrack Memory (RM), throughout the memory hierarchy. Thus, I developed FusedCache, a naturally inclusive, dual-level private cache design for RM that provides fast uniform access at one level, and non-uniform access at the next, which allows RM to be effective as close to the processor as an L1 cache. For higher levels of the memory hierarchy such as the last level cache, I propose a Multilane Racetrack Cache (MRC), an RM last level cache design utilizing lightweight compression combined with independent shifting. MRCs allow cache lines mapped to the same Racetrack structure to be accessed in parallel when compressed, mitigating potential shifting stalls in an RM cache. Finally, leveraging the lightweight compression from MRC and the need for efficient communication in Lynx, I present a cross-level design combining memory-level lightweight compression with network-level packet transfer, together with a technique called Source-Aware Layout Reorganization (SALR) to increase the compressibility of sensor data.