OATAO - Open Archive Toulouse Archive Ouverte Open Access Week

Hybrid direct and interactive solvers for sparse indefinite and overdetermined systems on future exascale architectures

Leleux, Philippe. Hybrid direct and interactive solvers for sparse indefinite and overdetermined systems on future exascale architectures. PhD, Mathématiques Appliquées, Institut National Polytechnique de Toulouse, 2021

[img]
Preview
(Document in English)

PDF (Author's version) - Requires a PDF viewer such as GSview, Xpdf or Adobe Acrobat Reader
6MB

Abstract

In scientific computing, the numerical simulation of systems is crucial to get a deep understanding of the physics underlying real world applications. The models used in simulation are often based on partial differential equations (PDE) which, after fine discretisation, give rise to huge sparse systems of equations to solve. Historically, 2 classes of methods were designed for the solution of such systems: direct methods, robust but expensive in both computations and memory; and iterative methods, cheap but with a very problem-dependent convergence properties. In the context of high performance computing, hybrid direct-iterative methods were then introduced inorder to combine the advantages of both methods, while using efficiently the increasingly largeand fast supercomputing facilities. In this thesis, we focus on the latter type of methods with two complementary research axis.In the first chapter, we detail the mechanisms behind the efficient implementation of multigrid methods. The latter makes use of several levels of increasingly refined grids to solve linear systems with a combination of fine grid smoothing and coarse grid corrections. The efficient parallel implementation of such a scheme is a difficult task. We focus on the solution of the problem on the coarse grid whose scalability is often observed as limiting at very large scales. We propose an agglomeration technique to gather the data of the coarse grid problem on a subset ofthe computing resources in order to minimise the execution time of a direct solver. Combined with a relaxation of the solution accuracy, we demonstrate an increased overall scalability of the multigrid scheme when using our approach compared to classical iterative methods, when the problem is numerically difficult. At extreme scale, this study is carried in the HHG framework(Hierarchical Hybrid Grids) for the solution of a Stokes problem with jumping coefficients, inspired from Earth's mantle convection simulation. The direct solver used on the coarse grid is MUMPS,combined with block low-rank approximation and single precision arithmetic.In the following chapters, we study some hybrid methods derived from the classical row-projection method block Cimmino, and interpreted as domain decomposition methods. These methods are based on the partitioning of the matrix into blocks of rows. Due to its known slow convergence, the original iterative scheme is accelerated with a stabilised block version of the conjugate gradient algorithm. While an optimal choice of block size improves the efficiency of this approach, the convergence stays problem dependent. An alternative solution is then introduced which enforces a convergence in one iteration by embedding the linear system into a carefully augmented space.These two approaches are extended in order to compute the minimum norm solution of in definite systems and the solution of least-squares problems. The latter problems require a partitioning in blocks of columns. We show how to improve the numerical properties of the iterative and pseudo-direct methods with scaling, partitioning and better augmentation methods. Both methods are implemented in the parallel solver ABCD-Solver (Augmented Block Cimmino Distributed solver)whose parallelisation we improve through a combination of load balancing and communication minimising techniques.Finally, for the solution of discretised PDE problems, we propose a new approach which augments the linear system using a coarse representation of the space. The size of the augmentation is controlled by the choice of a more or less refined mesh. We obtain an iterative method with fast linear convergence demonstrated on Helmholtz and Convection-Diffusion problems. The central point of the approach is the iterative construction and solution of a Schur complement

Item Type:PhD Thesis
Uncontrolled Keywords:
Institution:Université de Toulouse > Institut National Polytechnique de Toulouse - Toulouse INP (FRANCE)
Laboratory name:
Research Director:
Dayde, Michel and Ruiz, Daniel
Statistics:download
Deposited On:02 Sep 2021 12:56

Repository Staff Only: item control page