Skip to Main content Skip to Navigation

Athapascan-1 : interprétation distribuée du flot de données d'un programme parallèle

Abstract : The topic of this thesis is the modelisation by a data-flow graph of any execution of a parallel application. This graph, that links tasks and data, is dynamically built. This construction is independant from the effective tasks' scheduling. This independance enables the definition of a data access semantic and the control of memory consumption. We study in the first part the distributed algorithms enabling the construction and the management of this kind of graph. The central point of this management is the detection of the end of access of tasks on shared data. We propose a reactive algorithm performing this detection efficiently. The implementation of this algorithm is the kernel of the distributed implementation of the Athapascan-1 interface for parallel programming. The semantic of data access in this programming interface is lexicographic and its definition is based on the data-flow graph of the application. We show in a second part that the knowledge of the data-flow of an application enables a theoretical bound of the time and space of any execution. We propose, implement in Athapascan-1 and evaluate two distributed sheduling algorithms that limit the memory space used by any parallel execution. These experiments validate the theoretical results claimed by the two policies.
Complete list of metadatas

Cited literature [31 references]  Display  Hide  Download
Contributor : Thèses Imag <>
Submitted on : Wednesday, February 18, 2004 - 2:46:27 PM
Last modification on : Friday, November 6, 2020 - 4:39:03 AM
Long-term archiving on: : Friday, April 2, 2010 - 7:31:18 PM


  • HAL Id : tel-00004832, version 1




François Galilée. Athapascan-1 : interprétation distribuée du flot de données d'un programme parallèle. Réseaux et télécommunications [cs.NI]. Institut National Polytechnique de Grenoble - INPG, 1999. Français. ⟨tel-00004832⟩



Record views


Files downloads