Going beyond mean and median programs performances

Abstract : In the area of code performance optimisation and tuning, we are faced on the difficult problem of selecting the " best " code version based on empirical experiments and statistical analysis. With the massive introduction of general purpose multicore processors, programs performances become more and more instable, especially parallel programs. Usual statistical methods for computing performance speedups and comparing between programs are based on testing mean or median values. In this article, we explain why these metrics may be inadequate for making relevent decisions, and we propose new performance metrics based on parametric statistics using gaussian mixture models. Our new statistical methods are more accurate for decision making, they are formally defined, computed, implemented and distributed as free software in [1].
Type de document :
Communication dans un congrès
IEEE MCSoC 2016 : IEEE 10th International Symposium on Embedded Multicore/Many-core Systems-on-Chip, Sep 2016, Lyon, France
Liste complète des métadonnées

https://hal.inria.fr/hal-01371467
Contributeur : Sid Touati <>
Soumis le : lundi 26 septembre 2016 - 09:36:44
Dernière modification le : samedi 18 février 2017 - 01:20:38
Document(s) archivé(s) le : mardi 27 décembre 2016 - 12:20:49

Fichier

MCSoC-16-Worms-Touati.pdf
Fichiers produits par l'(les) auteur(s)

Identifiants

  • HAL Id : hal-01371467, version 1

Citation

Julien Worms, Sid Touati. Going beyond mean and median programs performances. IEEE MCSoC 2016 : IEEE 10th International Symposium on Embedded Multicore/Many-core Systems-on-Chip, Sep 2016, Lyon, France. <hal-01371467>

Partager

Métriques

Consultations de
la notice

122

Téléchargements du document

40