Ordonnancement de graphe dynamique de tâches sur architecture de grande taille

Abstract : In this phd, we focus on the efficient implementation of scheduling algorithms of actual parallel applications in high level parallel programing language. We propose mecanisms based on the principles of sequential and parallel degeneration. Sequential degeneration consist in optimising the cost of task creation when creation of more parallelism is not necessary. Distributed degeneration consist in the automatic generation of a distributed execution as close as possible to the same program written with a communication library as MPI. In practice, existing sequential degeneration mecanism achieve poor performances when implemented on distributed memory architecture. In ordre to couple these mecanisms, we have compared in term of page fault existing coherence protocoles permiting to implement distributed shared memory . This work permit to conclude that the "data-flow" coherence protocole proposed in this phd, permit to obtain good performances over distributed architecture when few tasks of the program are moved to another processeur. A distributed stack mecanism is then proposed in order to implement this protocole. Two scheduling algorithms, a static one and a dynamic one, permitting to exploit efficiently these mecanisms are also presented and evaluated. Finally, these mecanisms have been implemented in the Athapascan language and experimentally evaluated for simulation applications and for combinatorial optimisation applications.
Complete list of metadatas

https://tel.archives-ouvertes.fr/tel-00010909
Contributor : Rémi Revire <>
Submitted on : Tuesday, November 8, 2005 - 12:34:31 PM
Last modification on : Tuesday, April 24, 2018 - 1:30:28 PM
Long-term archiving on : Friday, April 2, 2010 - 10:21:27 PM

Identifiers

  • HAL Id : tel-00010909, version 1

Collections

INRIA | IMAG | UGA

Citation

Rémi Revire. Ordonnancement de graphe dynamique de tâches sur architecture de grande taille. Réseaux et télécommunications [cs.NI]. Institut National Polytechnique de Grenoble - INPG, 2004. Français. ⟨tel-00010909⟩

Share

Metrics

Record views

227

Files downloads

144