Hardware Awareness for the Selection of Optimal Iterative Linear Solvers

摘要

Solving sparse systems of linear equations is a commonly encountered computation in scientific and high-performance computing applications. Applications that depend on solving sparse linear systems as part of their workflow can spend a large percentage of their total runtime solving sparse systems. However, selecting the best iterative solver and preconditioner for solving a given sparse linear system, especially for novice users, is not a simple task. To address this problem, previous works have used machine learning techniques to find similarities between sparse matrices and the corresponding performance that solver-preconditioner pairs have on solving the resulting linear systems. This dissertation expands on existing work by introducing new techniques that incorporate hardware information into the prediction of ideal iterative linear solver and preconditioners for sparse linear systems. By accounting for hardware, it is possible to create more specially tailored solver-preconditioner recommendations for a novice user.