Skip to Main content Skip to Navigation

LaPIe : Communications Collectives adaptées aux Grilles de Calcul

Luiz Angelo Barchet-Estefanel 1
1 MOAIS - PrograMming and scheduling design fOr Applications in Interactive Simulation
ID-IMAG - Informatique et Distribution, Inria Grenoble - Rhône-Alpes, UJF - Université Joseph Fourier - Grenoble 1
Abstract : The popularity of heterogeneous parallel processing environments like clusters and computer grids has emphasised the impact of network heterogeneity on the performance of parallel applications. Collective communication operations are especially concerned by this problem, as communication heterogeneity interfers directly on the performance of the communication strategies. Therefore, the aim of this dissertation is the study the optimisation of collective communications in heterogeneous systems like computational grids. We investigate how to optimise communications according to the approach proposed by Karonis, where the communication among the different clusters that compose a grid are organised in multiple levels, minimising the total communication time (makespan). Hence, in a first moment, we study how to accurately model the performance of collective communications in homogeneous clusters, and, we validate such models under different network architectures. In parallel, we study how to minimise the impact of network heterogeneity on the complexity of communication scheduling techniques. Therefore, we investigated how topology discovery techniques can be used to decompose the grid environment into islands of homogeneous clusters, more adapted to an hierarchical modelling of the network. Finally, we study how performance modelling and topology discovery can be associated to construct gridaware collective communications. Using the information from both performance models and network topology, we compare different communication scheduling heuristics present on the literature. We also propose new heuristics to schedule inter-cluster communications, taking into account the performance of each different cluster. These heuristics seem to be specially efficient when the number of interconnected clusters augments, a tendency for the next years.
Complete list of metadatas
Contributor : Luiz Angelo Steffenel <>
Submitted on : Tuesday, February 14, 2006 - 11:39:49 AM
Last modification on : Friday, November 6, 2020 - 4:07:32 AM
Long-term archiving on: : Monday, September 17, 2012 - 11:55:36 AM


  • HAL Id : tel-00011603, version 1



Luiz Angelo Barchet-Estefanel. LaPIe : Communications Collectives adaptées aux Grilles de Calcul. Réseaux et télécommunications [cs.NI]. Institut National Polytechnique de Grenoble - INPG, 2005. Français. ⟨tel-00011603⟩



Record views


Files downloads