This thesis will include a collaboration with Jack Dongarra (University of Tennessee) and the company Metascale.
Abstract
One of the main challenges for the high-performance computing (HPC) community is to be able to propose software that is as much as possible independent from the hardware architecture. At the same time, HPC scientific applications address ever larger simulations in terms of amount of data or complexity of operations. As a result, the codes used in such simulations are not always affordable for scientists who are not expert in parallel programming. The scientific problem that will be considered at first is the solution of linear systems, which is at the heart of many HPC applications.
This PhD thesis will study automatic generation approaches to generate codes that take advantage of the features of the targeted architecture and of the intrinsic properties of the mathematical problem. These techniques will be also extended to linear-least squares and eigenvalue problems.
Context
Objectives
The objective of this PhD thesis is to propose a generic description of the linear problem to solve in order to exploit numerical and structure properties of matrices to get a fast and accurate solutions with respect to the type of problem. Information about targeted architecture and resources available will be also taken into account so that the most appropriate routine should be used or generated. A straightforward application of this generative approach is the possibility of prototyping new algorithms or new implementation of existing algorithms for various hardware. The software developed during this PhD thesis will be included into a generic scientific library that will be built on top of most recent numerical libraries for multicore and GPU architectures (e.g., MAGMA and PLASMA).
Work program
Extra information
Prerequisite
Détails
Expected funding
Institutional funding
Status of funding
Expected
Candidates
Utilisateur
marc.baboulin
Créé
Mardi 09 avril 2013 15:41:43 CEST
dernière modif.
Lundi 13 mai 2013 07:29:05 CEST
Fichiers joints
filename
créé
hits
filesize
Aucun fichier joint à cette fiche
Connexion
Ecole Doctorale Informatique Paris-Sud
Directrice
Nicole Bidoit Assistante
Stéphanie Druetta Conseiller aux thèses
Dominique Gouyou-Beauchamps
ED 427 - Université Paris-Sud
UFR Sciences Orsay
Bat 650 - aile nord - 417
Tel : 01 69 15 63 19
Fax : 01 69 15 63 87
courriel: ed-info à lri.fr