Performance predictor for HPF compilers
Dongsoo Han, Man-Pyo Hong · 2002
In HPF programming environment, program behavior needs to be reported for the development of well-tuned application programs. Performance predictor has been regarded as a proper approach for such requirement. In this paper, we propose a performance predictor which includes profiler, static performance predictor and dynamic performance predictor. Profiler generates execution count of each statement by profiling run of the program. While the static performance predictor generates general performance data, dynamic performance predictor reports some advanced performance data such as computation/communication overlapping and performance degradation by cache mis hit. Program slicing technique is applied to profiler and skeleton program is extracted from source program in dynamic performance predictor to speed up prediction time.