The Impact of High-Performance Computing on Linear Solver Algorithms and Mathematical Software

Industrial Problems Seminar

Ulrike Meier Yang
LLNL

Abstract

Linear Solvers are important mathematical algorithms needed in many applications. Often the linear systems that need to be solved are generated from the discretization of various differential equations. I have worked on various aspects of achieving good performance for linear solvers on ever evolving computer architectures with different features, increasing performance and memory capacities. Achieving good performance often requires significant changes to the software. Often, the best approach to reach this goal is the design of a new algorithm. I will share some of my experiences throughout my career from my early experiences as a student up to the time as lead of a project within DOE’s Exascale Computing Project. I will discuss some of the algorithms, and their implementations into software, such as the hypre library, and present performance results of the use of the software within an application on an exascale computer.