Perfect Sampling of Load Sharing Policies in Large Scale Distributed Systems

Gaël Gorgo 1 Jean-Marc Vincent 1
1 MESCAL - Middleware efficiently scalable
Inria Grenoble - Rhône-Alpes, LIG - Laboratoire d'Informatique de Grenoble
Abstract : This article presents a performance evaluation method for the dimensioning of load sharing policies in high performance distributed systems such as clusters and grids. Even for moderate system size, the corresponding Markovian models are not tractable neither analytically nor numerically.We propose a modelling framework and a simulation kernel which provides an unbiased sampling of the stationary distribution. As needed by the Propp & Wilson algorithm, we prove that events of load sharing systems preserve partial ordering on the state space (monotone events) that guarantees the simulation efficiency. This has been tested on large scale models (about 1000 nodes) in the psi2 simulation framework and applied for the comparison between work sharing and work stealing policies performances and for the optimisation of parameters such as the control rate and the probing depth.
Type de document :
Communication dans un congrès
ASMTA, LNCS, 2010, Cardiff, United Kingdom. Springer, 6148, pp.174-188, 2010, 〈10.1007/978-3-642-13568-2_13〉
Liste complète des métadonnées

https://hal.inria.fr/hal-00788885
Contributeur : Arnaud Legrand <>
Soumis le : vendredi 15 février 2013 - 13:11:21
Dernière modification le : mercredi 14 décembre 2016 - 01:08:43

Identifiants

Collections

Citation

Gaël Gorgo, Jean-Marc Vincent. Perfect Sampling of Load Sharing Policies in Large Scale Distributed Systems. ASMTA, LNCS, 2010, Cardiff, United Kingdom. Springer, 6148, pp.174-188, 2010, 〈10.1007/978-3-642-13568-2_13〉. 〈hal-00788885〉

Partager

Métriques

Consultations de la notice

136