Section: New Results

Parametric and Non Parametric Statistics for Code Performance Analysis

Participant : Sid Touati.

This activity is conducted by Sid Touati in collaboration with Julien Worms, an associate professor in Mathematics at the university of Versailles Saint Quentin. It was started under the consideration that the performances of programs are hardly ever represented by a gaussian distribution. So, our purpose here is to study parametric statistics for analyzing the performances of programs. We are interested in modelling program performances by gaussian mixtures (using mixmod method). After a statistical modeling, we deduce multiple performance tests and performance criteria to decide with a high degree of confidence about the "best" program run version. This is still work-in-progress: we are implementing a free software for analysis based on our approach, and we are writing a rather complete research report prior to further publications in conferences and journals.