Soft real-time embedded systems, such as PDAs, portable multimedia devices, sensors, and some control systems, require the processing of data streams in a timely fashion, as determined by the satisfaction of end users who cannot detect occasional execution failures or deadline misses. Additional system resources would provide better performance, but the end users may not notice and will not reward such efforts. This unique feature sheds light on the efficient design and implementation of soft real-time systems. This project is developing a new design framework to systematically transfer such softness of the deadlines into better resource management, such as reduction of energy consumption. The strategy pursued in this research is to allocate (off-line) the minimal system resources necessary to meet the given soft real-time performance requirements statistically. That is, the system does not guarantee the completion of each execution, but it will produce sufficiently many completions over a large number of repetitions to meet the user-specific completion ratio. Then during real time execution, on-chip temperature on system speed and power dissipation guide dynamic scaling of frequency and voltage to exploit unused capacity for energy efficiency on the multi-processor system. This work focuses on (1) building and validating a theoretical foundation for the proposed probabilistic design framework; (2) developing a set of probabilistic multimedia application benchmarks and prototype synthesis tool software; and (3) conducting proof-of-concept prototyping of consumer-oriented digital signal processing software. This research enables the resource-efficient design and implementation of an important family of soft real-time embedded systems and is expected to have direct impact on the embedded systems industry, future surveillance systems, and people's daily life. Both undergraduate and graduate students involved in the project learn embedded system design skills and gain research experience.
