Solverstack@Inria Bordeaux

Solverstack@Inria Bordeaux aims at providing a high-performance (HPC) linear algebra solver stack. It provides a collection of solvers, partitioning tools and runtimes systems, which may be employed on modern supercomputers to operate on dense and sparse matrices. It includes both direct and iterative methods (together with various preconditioners), as well as hybrid direct/iterative ones.

Solverstack DAG

One of the main objective is the portability of performance, from the laptops to the supercomputer. The software are mainly developed in C, C++ and Fortran by researchers and engineers working in -- or in strong collababoration with -- Inria joint project-teams with Bordeaux INP, Bordeaux University and CNRS (LaBRI UMR 5800) HiePACS, Storm and TADaaM located at Inria Bordeaux Sud-Ouest.

Solverstack is a collection of software libraries, with some dependencies between them. This software collection provides:

  • dense direct methods in Chameleon,
  • dense matrix randomized eigen and singular value decomposition methods in Fmr,
  • multivariate analysis (MVA) methods in Cppdiodon,
  • sparse direct methods in PaStiX and qr_mumps,
  • Krylov subspace iterative methods in fabulous,
  • hybrid direct/iterative methods in MaPHyS++,
  • a task-based support with the StarPU runtime system,
  • graphs partitioning and sparse matrix ordering with Scotch.

See libraries features.

Check the news to get latest releases.

Solverstack is packaged with care for the high-performance guix-hpc and spack (scotch, starpu, chameleon and pastix are available in the official spack repository) deployment tools as well as the debian/ubuntu and brew distributions (scotch is available in the official brew repository).