This project aims to develop a framework for performance characterization, representation, optimization and debugging for the emerging domain of soft real time computing. Key elements include a new metric, called "variance", that characterizes the timing properties of such workloads, and a "variant characterization graph" (VCG) that represents the relationships among the variance contributing functions in a software system. The VCG is subjected to a combination of static analysis, dynamic profiling and statistical analysis. The results of these analyses are applied to the problems of debugging performance bottlenecks, and dynamic performance optimization making use of predictive methods and contextual information obtained at run time.
Interactive, response-time sensitive applications such as video games, computer vision, and multi-media encoder/decoders are an important, growing class of emerging applications on mobile devices such as the smart-phones. A proliferation of powerful new devices is driving up the scale and pervasiveness of applications. The diversity of new devices and the number of kinds of resources present in them, accompanied by a market-driven demand for shorter application development cycles, are making performance tuning more challenging and more essential. Success of this project would provide a practical conceptual backbone for converting the optimization of such soft real time systems from an art to a science. Transition of the results of this research to practice is furthered through open source distribution of software tools implementing and applying the VCG-based analyses, and an on-line graduate course titled "Optimizations for Interactive Embedded Software".
