Déploiement et contrôle d'applications parallèles sur grappes de grandes tailles

• Main Author: Martin, Cyrille
• Other Authors: Informatique et Distribution (ID-IMAG), Institut National de Recherche en Informatique et en Automatique (Inria)-Institut National Polytechnique de Grenoble (INPG)-Centre National de la Recherche Scientifique (CNRS), Institut National Polytechnique de Grenoble - INPG, Plateau Brigitte
• Format: Doctoral or Postdoctoral Thesis
• Language: French
• Published: HAL CCSD 2003
• Subjects: deployment; parallel launch; broadcast; scheduling; remote execution call protocol; scalability; efficient; generic; déploiement; diffusion; ordonnancement; protocole d'exécution distante; efficace; générique; passage à l'échelle; [INFO.INFO-NI]Computer Science [cs]/Networking and Internet Architecture [cs.NI]; Rosse; inuktitut
• Online Access: https://theses.hal.science/tel-00004610; https://theses.hal.science/tel-00004610/document; https://theses.hal.science/tel-00004610/file/tel-000046101.pdf

M. Jacques Mossière Président M. Franck Cappello Rapporteur M. Raymond Namyst Rapporteur Mme Brigitte Plateau Examinateur (Directeur de thèse) M. Jacques Briat Examinateur (Co-encadrant) Mme Pascale Rosse Examinateur (BULL sa) The increasing size of cluster of workstations sets down the scalability problem of applications running on these platforms. This concerns both numerical parallel applications and exploitation tools (administration, monitoring.). In this thesis work, we study the deployment of parallel applications on large clusters, that can be extended to grids. The deployment includes on one hand the launch of the parallel program on all nodes and on the other hand the setting up of a communication layer. Efficiency is obtained thanks to the overlay of all independent steps of the deployment. This work shows this problem as equivalent as the well known problem of the single message broadcast. Performance gap between the cost of a network communication and this of a remote execution call enable us to use a work stealing algorithm to realize a near-optimal schedule of remote execution calls. The good properties and performance figures of this tool, Taktuk, are demonstrated by its use in several projects like: KaTools (included and used by the Clic Mandrake Cluster Linux distribution), OAR (Job manager) and Inuktitut (Communication layer of the environment ATHAPASCAN). La taille grandissante des grappes de calcul pose le problème du "passage à l'échelle" des applications qui s'exécutent sur ces plates-formes. Ceci concerne les applications de calculs scientifiques et les applications permettant d'exploiter ces plates-formes (administration, surveillance de charge, etc.). Dans ce travail de thèse nous nous sommes intéressés au déploiement d'une application parallèle sur une grappe de grande taille. L'objectif de cette étude était de fournir une méthode de déploiement efficace sur des grappes composées de milliers de noeuds et pouvant être facilement étendue aux grilles de calcul. Le déploiement inclut ...