Computer Science Research Seminar

Wednesday, November 23rd, 2005

Alex Doboli

Electrical and Computer Engineering Department
Stony Brook University

Title: Middleware techniques for on-line adaptation of embedded architectures


Sensor networks are emerging as a key technology for many applications in national security, health care, environmental monitoring, infrastructure security, food safety, manufacturing automation and many more. A key problem is that of devising embedded software and hardware architectures that can effectively operate in continuously changing, hard-to-predict conditions. In addition, architectures should be cheap and consume tiny amounts of energy, considering that their batteries are hard to be replaced or replenished. Varying throughput constraints necessitate a continuous adaptation in an embedded system's architecture, including selection of the number of powered up resources, the supply voltages and the clock frequencies. This talk presents a novel methodology, which extracts an optimal set of design points for an adaptable embedded system under continuously varying throughput constraints while minimizing the resulting system over-design. We discuss online control methods which adapt the architecture to minimize energy consumption while satisfying varying latency constraints. Adaptive image processing for vehicle tracking is used as a case study for these approaches.