Design Methods for High-Performance Sensor Networks

Princeton University, Princeton NJ

Investigators

Abstract

Wolf Abstract This research is developing new methods for the design of high-performance sensor networks that perform significant amounts of computation within the sensor network. High- bandwidth signal processing applications, such as distributed audio or video processing, are becoming increasingly common, and performing large amounts of processing in the distributed system that implements the sensor network poses new design challenges. On the one hand, high-performance sensor networks are like distributed embedded systems, which need to be optimized to the application in order to minimize power consumption and cost and maximize performance; on the other hand, they are like ad-hoc networks that need to be able to be operate in a number of different configurations without radically increasing the cost of operating the network. New design methodologies and tools are needed that can create a resilient architecture that still takes advantage of application-specific characteristics. Traditional algorithms and methodologies for designing distributed embedded systems can handle some of the aspects of sensor networks, namely distributed computation and the effects of communication delay, but their purpose is to select a single good design. They provide no guarantees that any changes in the system parameters (number of nodes, communication, power budget) will result in a system that operates anywhere close to optimality. With the proper choice of nodes and links, and the right topology and communication patterns within the high-performance sensor network, the cost/power/performance of the network can be substantially improved. A combination of design-time and run-time decisions is required for the successful design and deployment of high-performance sensor networks because certain characteristics of the system will not be known until run time, and those characteristics may change during operation of the network. Furthermore, the configuration of the network may not be known at design time. The overall hardware and software architecture of the network must be designed to operate well not just at a single design point, but across a range of possible configurations and operating decisions. This research seeks to balance the needs for optimizing the design and adapting its operation at run time by developing new design methods that find pareto-optimal regions in the design space. Pareto-optimality has been used to design small-scale embedded systems, but this work improves on previous efforts in several ways: handling larger systems, lower-variance design; analysis of run-time behavior; and inclusion of Monte- Carlo methods. Broader impacts: The methods developed in the research will directly enable systems for security and a variety of industrial applications. High-performance audio and video are especially important in security applications and distributed real-time processing of data will help to create security systems that can identify problems more quickly. Design tools, benchmark data, and lecture materials on distributed embedded systems will be distributed over the web.