Section: Research Program

High Performance Scientific Computing

This research is in the area of high performance scientific computing, and in particular in parallel matrix algorithms. This is a subject of crucial importance for numerical simulations as well as other scientific and industrial applications, in which linear algebra problems arise frequently. The modern numerical simulations coupled with ever growing and more powerful computational platforms have been a major driving force behind a progress in numerous areas as different as fundamental science, technical/technological applications, life sciences.

The main focus of this research is on the design of efficient, portable linear algebra algorithms, such that solving a large set of linear equations or a least squares problem. The characteristics of the matrices commonly encountered in this situations can vary significantly, as are the computational platforms used for the calculations. Nonetheless two common trends are easily discernible. First, the problems to solve are larger and larger, since the numerical simulations are using higher resolution. Second, the architecture of today's supercomputers is getting very complex, and so the developed algorithms need to be adapted to these new achitectures.

Communication avoiding algorithms for numerical linear algebra

Since 2007, we work on a novel approach to dense and sparse linear algebra algorithms, which aims at minimizing the communication, in terms of both its volume and a number of transferred messages. This research is motivated by technological trends showing an increasing communication cost. Its main goal is to reformulate and redesign linear algebra algorithms so that they are optimal in an amount of the communication they perform, while retaining the numerical stability. The work here involves both theoretical investigation and practical coding on diverse computational platforms. We refer to the new algorithms as communication avoiding algorithms  [58]  [10] . In our team we focus on communication avoiding algorithms for dense direct methods as well as sparse iterative methods.

The theoretical investigation focuses on identifying lower bounds on communication for different operations in linear algebra, where communication refers to data movement between processors in the parallel case, and to data movement between different levels of memory hierarchy in the sequential case. The lower bounds are used to study the existing algorithms, understand their communication bottlenecks, and design new algorithms that attain them.

This research focuses on the design of linear algebra algorithms that minimize the cost of communication. Communication costs include both latency and bandwidth, whether between processors on a parallel computer or between memory hierarchy levels on a sequential machine. The stability of the new algorithms represents an important part of this work.

Preconditioning techniques

Solving a sparse linear system of equations is the most time consuming operation at the heart of many scientific applications, and therefore it has received a lot of attention over the years. While direct methods are robust, they are often prohibitive because of their time and memory requirements. Iterative methods are widely used because of their limited memory requirements, but they need an efficient preconditioner to accelerate their convergence. In this direction of research we focus on preconditioning techniques for solving large sparse systems.

One of the main challenges that we address is the scalability of existing methods as incomplete LU factorizations or Schwarz-based approaches, for which the number of iterations increases significantly with the problem size or with the number of processors. This is often due to the presence of several low frequency modes that hinder the convergence of the iterative method. To address this problem, we study direction preserving solvers in the context of multilevel filtering LU decompositions. A judicious choice for the directions to be preserved through filtering allows us to alleviate the effect of low frequency modes on the convergence. While preconditioners and their scalability are studied by many other groups, our approach of direction preserving and filtering is studied in only very few other groups in the world (as Lawrence Livermore National Laboratory, Frankfurt University, Pennsylvania State University).

Fast linear algebra solvers based on randomization

Linear algebra calculations can be enhanced by statistical techniques in the case of a square linear system $Ax=b$ where $A$ is a general or symmetric indefinite matrix  [3] & [1] . Thanks to a random transformation of $A$, it is possible to avoid pivoting and then to reduce the amount of communication. Numerical experiments show that this randomization can be performed at a very affordable computational price while providing us with a satisfying accuracy when compared to partial pivoting. This random transformation called Partial Random Butterfly Transformation (PRBT) is optimized in terms of data storage and flops count. A PRBT solver for LU factorization (and for $LD{L}^T$ factorization on multicore) has been developed. This solver takes advantage of the latest generation of hybrid multicore/GPU machines and gives better Gflop/s performance than existing factorization routines [19] .

Sensitivity analysis of linear algebra problems

We derive closed formulas for the condition number of a linear function of the total least squares solution [4] . Given an over determined linear systems $Ax=b$, we show that this condition number can be computed using the singular values and the right singular vectors of $[A,b]$ and $A$. We also provide an upper bound that requires the computation of the largest and the smallest singular value of $[A,b]$ and the smallest singular value of $A$. In numerical experiments, we compare these values with condition estimates from the literature.