Skip to Main content Skip to Navigation
Theses

Placement de taches sur ordinateurs paralleles a memoire distribuee

Abstract : The growing needs in computing performance imply more complex computer architectures. The lack of good programming environments for these machines must be filled. The goal to be reached is to find a compromise solution between portability and performance. The subject of this thesis is studying the problem of static allocation of task graphs onto distributed memory parallel computers. This work takes part of the project INRIA-IMAG APACHE and of the european one SEPP-COPERNICUS (Software Engineering for Parallel Processing). The undirected task graph is the chosen programming model. A survey of the existing solutions for scheduling and for mapping problems is given. The possibility of using directed task graphs after a clustering phase is underlined. An original solution is designed and implemented ; this solution is implemented within a working programming environment. Three kinds of mapping algorithms are used: greedy, iterative and exact ones. Most developments have been done for tabu search and simulated annealing. These algorithms improve various objective functions (from most simple and portable to the most complex and architecturaly dependant). The weigths of the task graphs can be tuned using a post-mortem analysis of traces. The use of tracing tools leads to a validation of the cost function and of the mapping algorithms. A benchmark protocol is defined and used. The tests are runned on the Meganode (a 128 transputer machine) using VCR from the university of Southampton as a router, synthetic task graphs generation with ANDES of the ALPES project (developped by the performance evaluation team of the LGI-IMAG) and the Dominant Sequence Clustering of PYRROS (developped by Tao Yang and Apostolos Gerasoulis).
Complete list of metadatas

Cited literature [21 references]  Display  Hide  Download

https://tel.archives-ouvertes.fr/tel-00005081
Contributor : Thèses Imag <>
Submitted on : Wednesday, February 25, 2004 - 11:20:11 AM
Last modification on : Friday, November 6, 2020 - 4:12:49 AM
Long-term archiving on: : Wednesday, September 12, 2012 - 4:55:24 PM

Identifiers

  • HAL Id : tel-00005081, version 1

Collections

CNRS | IMAG | UGA

Citation

Pascal Bouvry. Placement de taches sur ordinateurs paralleles a memoire distribuee. Réseaux et télécommunications [cs.NI]. Institut National Polytechnique de Grenoble - INPG, 1994. Français. ⟨tel-00005081⟩

Share

Metrics

Record views

373

Files downloads

390