π™Όπš„π™Όπ™Ώπš‚: MUltifrontal Massively Parallel Solver for the direct solution of sparse linear equations

  • Patrick R. Amestoy

    Mumps Technologies SAS, Lyon, France
  • Jean-Yves L’Excellent

    Mumps Technologies SAS, Lyon, France
  • Theo Mary

    Sorbonne UniversitΓ©, CNRS, LIP6, Paris, France
  • Chiara Puglisi

    Mumps Technologies SAS, Lyon, France
π™Όπš„π™Όπ™Ώπš‚: MUltifrontal Massively Parallel Solver for the direct solution of sparse linear equations cover
Download Chapter PDF

This book chapter is published open access.

Abstract

The scientific library π™Όπš„π™Όπ™Ώπš‚ (MUltifrontal Massively Parallel Solver) solves large systems of sparse linear equations using a direct method. It is an open-source software library distributed under the CeCILL-C license which can be downloaded from the π™Όπš„π™Όπ™Ώπš‚ solver website.

Direct methods are widely appreciated for their numerical robustness, reliability, and ease of use. They are commonly used either to compute the solution or as a preconditioner within iterative methods. However, they are also known for their high computational complexity.

In this communication, we briefly describe the main features of the π™Όπš„π™Όπ™Ώπš‚ solver. We then explain how data sparsity with block low-rank (BLR) approximations and mixed precision can be used to reduce the complexity of direct methods. We conclude by showing results on large real-world applications (matrices of order several hundred million) and describing ongoing research work.