Sujets

View

View Item

Domain Parallelism-High Performance Computing-Grid
Domain - extra HPC
Year 2014
Starting October 2014
Status Open
Subject Randomized algorithms in large scale matrix computations
Thesis advisor BABOULIN Marc
Co-advisors
Laboratory LRI ParSys
Collaborations University of Tennessee (Knoxville, USA), Lawrence Berkeley National Laboratory (Berkeley, USA)
Abstract The development of new parallel architectures and accelerators requires the rethinking of most of the numerical linear algebra algorithms which are at the heart of many scientific applications. As a result, innovative methods must be proposed in order to take full advantage of current supercomputers and among them is the use of randomized algorithms. In recent years, techniques based on Random Butterfly Transformations (RBT) have been successfully proposed to accelerate the solution of dense systems. The performance and accuracy results are encouraging on current hybrid multicore/GPU architectures. In this PhD thesis we propose to extend the use of randomized algorithms to sparse (direct and/or iterative) linear system solvers which concern a large class of physical and industrial applications and represent a challenging issue for future exascale systems.
Context This PhD thesis will take advantage of the new associate-team R-LAS with University of Tennessee (project lead by Marc Baboulin and Jack Dongarra). In the framework of this project, the PhD candidate will benefit from extended visits to Innovative Computing Laboratory (Knoxville, Tennessee), which is one of the world leader in High-Performance Computing.
Objectives This PhD thesis aims at developing innovative randomized algorithms and software for sparse matrix computations, either for direct or iterative methods. The resulting software will be applied to large-size simulations addressed by ParSys and its Paris-Saclay partners in the area of high-performance computing (LIMSI, EDF, Centrale Paris). Moreover, for sake of visibility of the results, the software developed during this thesis would be part of a reference public domain library for fast HPC solvers.
Work program
Extra information
Prerequisite
Details
Expected funding Institutional funding
Status of funding Expected
Candidates
user marc.baboulin
Created Wednesday 26 of February, 2014 12:31:31 CET
LastModif Thursday 01 of May, 2014 15:46:28 CEST
Comments
Attachments (0)

Attachments

 filenamecreatedhitsfilesize 
No attachments for this item


The original document is available at https://edips.lri.fr/tiki-view_tracker_item.php?itemId=3744