Using static analysis data for performance modeling and prediction

In general, high performance computing applications have large codebases composed of various scientific algorithms which must be tuned to achieve optimal speed. Therefore, a programmer extracts pieces of code from large programs, as candidates for the performance tuning. Maximizing such code performance requires measurement, analysis, and optimization strategies, targeting hardware components. Furthermore, computer architecture improvement raises hardware co-design issues such as measuring detailed computer performance. Currently, code execution time is well measured, but it is much harder to break out the performance contributory details per hardware resource in order to predict a code performance. This paper presents the Ubenchface tool, a framework for performance prediction and knowledge discovery. Inversely to traditional measurement methods and modeling, the proposed tool considers static metrics to analyze and tune application performance. This framework is more informative than simple benchmarking, or microbenchmarking. It is useful for performance investigations in similarity and redundancy study concerning benchmark suites, predicting, understanding scaling, and tuning.