Skip to Main content Skip to Navigation

Minimisation du sur-coût des communications dans la parallélisation des algorithmes numériques

Abstract : The aim of this thesis is the study of different methods to minimize the communication overhead due to the parallelization of numerical kernels. The first method consists in optimizing collective communications algorithms. We have proposed novel algorithms to achieve matrix transpose, for squared matrices distributed in a block fashion. We have also studied the total exchange problem. This communication scheme is useful in the parallelization of numerical kernels (as, for instance, the conjugate gradient algorithm). We have proposed efficient algorithms of total echange for torus topologies. The second method consists in overlapping communications by computations. We have studied some basic algorithmic principles which allow the overlap of communications. These ones are based on pipelinig technics and local computational task reordering. These technics have been illustrated in the parallelization of Fourier transform algorithms. The different implementations of these algrithms on various distributed memory parallel machines (Cray T3D, IBM SP2, Intel iPSC-860 and Paragon) highlight the gain of efficiency induced using thes technics.
Complete list of metadata
Contributor : Thèses Imag <>
Submitted on : Tuesday, February 24, 2004 - 9:59:23 AM
Last modification on : Wednesday, March 10, 2021 - 1:50:03 PM
Long-term archiving on: : Friday, September 14, 2012 - 11:30:31 AM


  • HAL Id : tel-00005034, version 1




Christophe Calvin. Minimisation du sur-coût des communications dans la parallélisation des algorithmes numériques. Réseaux et télécommunications [cs.NI]. Institut National Polytechnique de Grenoble - INPG, 1995. Français. ⟨tel-00005034⟩



Record views


Files downloads