To leverage the extreme parallelism of emerging architectures, so that scientific applications can fulfill their high fidelity and multi-physics potential while sustaining high efficiency relative to the limiting resource, numerical algorithms must be redesigned. Algorithmic redesign is capable of shifting the limiting resource, for example from memory or communication to arithmetic capacity. The benefit of algorithmic redesign expands greatly when introducing a tunable tradeoff between accuracy and resources. Scientific applications from diverse sources rely on dense matrix operations. These operations arise in: Schur complements, integral equations, covariances in spatial statistics, ridge regression, radial basis functions from unstructured meshes, and kernel matrices from machine learning, among others. This thesis demonstrates how to extend the problem sizes that may be treated and to reduce their execution time. Two “universes” of algorithmic innovations have emerged to improve computations by orders of magnitude in capacity and runtime. Each introduces a hierarchy, of rank or precision. Tile Low-Rank approximation replaces blocks of dense operator with those of low rank. Mixed precision approximation, increasingly well supported by contemporary hardware, replaces blocks of high with low precision. Herein, we design new high-performance direct solvers based on the synergism of TLR and mixed precision. Since adapting to data sparsity leads to heterogeneous workloads, we rely on task-based runtime systems to orchestrate the scheduling of fine-grained kernels onto computational resources. We first demonstrate how TLR permits to accelerate acoustic scattering and mesh deformation simulations. Our solvers outperform the state-of-art libraries by up to an order of magnitude. Then, we demonstrate the impact of enabling mixed precision in bioinformatics context. Mixed precision enhances the performance up to three-fold speedup. To facilitate the adoption of task-based runtime systems, we introduce the AL4SAN library to provide a common API for the expression and queueing of tasks across multiple dynamic runtime systems. This library handles a variety of workloads at a low overhead, while increasing user productivity. AL4SAN enables interoperability by switching runtimes at runtime, which permits to achieve a twofold speedup on a task-based generalized symmetric eigenvalue solver.

This dissertation develops adaptive hp-Finite Element (FE) technology and a parallel sparse direct solver enabling the accurate modeling of the absorption of Electro-Magnetic (EM) energy in the human head. With a large and growing number of cell phone users, the adverse health effects of EM fields have raised public concerns. Most research that attempts to explain the relationship between exposure to EM fields and its harmful effects on the human body identifies temperature changes due to the EM energy as the dominant source of possible harm. The research presented here focuses on determining the temperature distribution within the human body exposed to EM fields with an emphasis on the human head. Major challenges in accurately determining the temperature changes lie in the dependence of EM material properties on the temperature. This leads to a formulation that couples the BioHeat Transfer (BHT) and Maxwell equations. The mathematical model is formed by the time-harmonic Maxwell equations weakly coupled with the transient BHT equation. This choice of equations reflects the relevant time scales. With a mobile device operating at a single frequency, EM fields arrive at a steady-state in the micro-second range. The heat sources induced by EM fields produce a transient temperature field converging to a steady-state distribution on a time scale ranging from seconds to minutes; this necessitates the transient formulation. Since the EM material properties depend upon the temperature, the equations are fully coupled; however, the coupling is realized weakly due to the different time scales for Maxwell and BHT equations. The BHT equation is discretized in time with a time step reflecting the thermal scales. After multiple time steps, the temperature field is used to determine the EM material properties and the time-harmonic Maxwell equations are solved. The resulting heat sources are recalculated and the process continued. Due to the weak coupling of the problems, the corresponding numerical models are established separately. The BHT equation is discretized with H¹ conforming elements, and Maxwell equations are discretized with H(curl) conforming elements. The complexity of the human head geometry naturally leads to the use of tetrahedral elements, which are commonly employed by unstructured mesh generators. The EM domain, including the head and a radiating source, is terminated by a Perfectly Matched Layer (PML), which is discretized with prismatic elements. The use of high order elements of different shapes and discretization types has motivated the development of a general 3D hp-FE code. In this work, we present new generic data structures and algorithms to perform adaptive local refinements on a hybrid mesh composed of different shaped elements. A variety of isotropic and anisotropic refinements that preserve conformity of discretization are designed. The refinement algorithms support one- irregular meshes with the constrained approximation technique. The algorithms are experimentally proven to be deadlock free. A second contribution of this dissertation lies with a new parallel sparse direct solver that targets linear systems arising from hp-FE methods. The new solver interfaces to the hierarchy of a locally refined mesh to build an elimination ordering for the factorization that reflects the h-refinements. By following mesh refinements, not only the computation of element matrices but also their factorization is restricted to new elements and their ancestors. The solver is parallelized by exploiting two-level task parallelism: tasks are first generated from a parallel post-order tree traversal on the assembly tree; next, those tasks are further refined by using algorithms-by-blocks to gain fine-grained parallelism. The resulting fine-grained tasks are asynchronously executed after their dependencies are analyzed. This approach effectively reduces scheduling overhead and increases flexibility to handle irregular tasks. The solver outperforms the conventional general sparse direct solver for a class of problems formulated by high order FEs. Finally, numerical results for a 3D coupled BHT with Maxwell equations are presented. The solutions of this Maxwell code have been verified using the analytic Mie series solutions. Starting with simple spherical geometry, parametric studies are conducted on realistic head models for a typical frequency band (900 MHz) of mobile phones. / text

Dans cette thèse de doctorat, nous étudions des algorithmes et des implémentations pour accélérer la résolution de systèmes linéaires denses en utilisant des architectures composées de processeurs multicœurs et d'accélérateurs. Nous nous concentrons sur des méthodes basées sur la factorisation LU. Le développement de notre code s'est fait dans le contexte de la bibliothèque MAGMA. Tout d'abord nous étudions différents solveurs CPU/GPU hybrides basés sur la factorisation LU. Ceux-ci visent à réduire le surcoût de communication dû au pivotage. Le premier est basé sur une stratégie de pivotage dite "communication avoiding" (CALU) alors que le deuxième utilise un préconditionnement aléatoire du système original pour éviter de pivoter (RBT). Nous montrons que ces deux méthodes surpassent le solveur utilisant la factorisation LU avec pivotage partiel quand elles sont utilisées sur des architectures hybrides multicœurs/GPUs. Ensuite nous développons des solveurs utilisant des techniques de randomisation appliquées sur des architectures hybrides utilisant des GPU Nvidia ou des coprocesseurs Intel Xeon Phi. Avec cette méthode, nous pouvons éviter l'important surcoût du pivotage tout en restant stable numériquement dans la plupart des cas. L'architecture hautement parallèle de ces accélérateurs nous permet d'effectuer la randomisation de notre système linéaire à un coût de calcul très faible par rapport à la durée de la factorisation. Finalement, nous étudions l'impact d'accès mémoire non uniformes (NUMA) sur la résolution de systèmes linéaires denses en utilisant un algorithme de factorisation LU. En particulier, nous illustrons comment un placement approprié des processus légers et des données sur une architecture NUMA peut améliorer les performances pour la factorisation du panel et accélérer de manière conséquente la factorisation LU globale. Nous montrons comment ces placements peuvent améliorer les performances quand ils sont appliqués à des solveurs hybrides multicœurs/GPU. / In this PhD thesis, we study algorithms and implementations to accelerate the solution of dense linear systems by using hybrid architectures with multicore processors and accelerators. We focus on methods based on the LU factorization and our code development takes place in the context of the MAGMA library. We study different hybrid CPU/GPU solvers based on the LU factorization which aim at reducing the communication overhead due to pivoting. The first one is based on a communication avoiding strategy of pivoting (CALU) while the second uses a random preconditioning of the original system to avoid pivoting (RBT). We show that both of these methods outperform the solver using LU factorization with partial pivoting when implemented on hybrid multicore/GPUs architectures. We also present new solvers based on randomization for hybrid architectures for Nvidia GPU or Intel Xeon Phi coprocessor. With this method, we can avoid the high cost of pivoting while remaining numerically stable in most cases. The highly parallel architecture of these accelerators allow us to perform the randomization of our linear system at a very low computational cost compared to the time of the factorization. Finally we investigate the impact of non-uniform memory accesses (NUMA) on the solution of dense general linear systems using an LU factorization algorithm. In particular we illustrate how an appropriate placement of the threads and data on a NUMA architecture can improve the performance of the panel factorization and consequently accelerate the global LU factorization. We show how these placements can improve the performance when applied to hybrid multicore/GPU solvers.