A unified methodology for a fast benchmarking of parallel architecture
Abstract
Benchmarking of architectures is today jeopardized by the explosion of parallel architectures and the dispersion of parallel programming models. Parallel programming requires architecture dependent compilers and languages as well as high programming expertise. Thus, an objective comparison has become a harder task. This paper presents a novel methodology to evaluate and to compare parallel architectures in order to ease the programmer work. It is based on the usage of microbenchmarks, code profiling and characterization tools. The main contribution of this methodology is a semi-automatic prediction of the performance for sequential applications on a set of parallel architectures. In addition the performance estimation is correlated with the cost of other criteria such as power or portability. Our methodology prediction was validated on an industrial application. Results are within a range of 20%.