Deskilling HPL using an evolutionary algorithm to automate cluster benchmarking

Abstract : The High-Performance Linpack (HPL) benchmark is the accepted standard for measuring the capacity of the world's most powerful computers, which are ranked twice yearly in the Top 500 List. Since just a small deficit in performance can cost a computer several places, it is important to tune the benchmark to obtain the best possible result. However, the adjustment of HPL's seventeen configuration parameters to obtain maximum performance is a time-consuming task that must be performed by hand. In a previous paper, we provided a preliminary study that proposed the tuning of HPL parameters by means of an Evolutionary Algorithm. The approach was validated on a small cluster. In this article, we extend this initial work by describing Acbea, a fullyautomatic benchmark tuning tool that performs both the configuration and installation of HPL followed by an automatic search for optimized parameters that will lead to the best benchmark results. Experiments have been conducted to validate this tool on several clusters, exploiting in particular the Grid'5000 infrastructure.
Type de document :
Communication dans un congrès
Springer. PPAM 2009 : 8th International Conference on Parallel Processing and Applied Mathematics, Sep 2009, Wroclaw, Poland. Springer, 6068, pp.102-114, 2010, LNCS. 〈10.1007/978-3-642-14403-5_12〉
Liste complète des métadonnées

https://hal.inria.fr/hal-00692022
Contributeur : Ist Rennes <>
Soumis le : vendredi 27 avril 2012 - 16:40:25
Dernière modification le : lundi 20 juin 2016 - 14:10:32

Identifiants

Collections

Citation

Dominic Dunlop, Sébastien Varrette, Pascal Bouvry. Deskilling HPL using an evolutionary algorithm to automate cluster benchmarking. Springer. PPAM 2009 : 8th International Conference on Parallel Processing and Applied Mathematics, Sep 2009, Wroclaw, Poland. Springer, 6068, pp.102-114, 2010, LNCS. 〈10.1007/978-3-642-14403-5_12〉. 〈hal-00692022〉

Partager

Métriques

Consultations de la notice

38