A generic library for large scale solution of PDEs on modern heterogeneous architectures

Stefan Lemvig Glimberg, Allan Peter Engsig-Karup

    Research output: Chapter in Book/Report/Conference proceedingConference abstract in proceedingsResearchpeer-review

    137 Downloads (Pure)

    Abstract

    Adapting to new programming models for modern multi- and many-core architectures requires code-rewriting and changing algorithms and data structures, in order to achieve good efficiency and scalability. We present a generic library for solving large scale partial differential equations (PDEs), capable of utilizing heterogeneous CPU/GPU environments. The library can be used for fast proto-typing of PDE solvers, based on finite difference approximations of spatial derivatives in one, two, or three dimensions. In order to efficiently solve large scale problems, we keep memory consumption and memory access low, using a low-storage implementation of flexible-order finite difference operators. We will illustrate the use of library components by assembling such matrix-free operators to be used with one of the supported iterative solvers, such as GMRES, CG, Multigrid or Defect Correction. As a proto-type for large scale PDE solvers, we present the assembling of a tool for simulation of three dimensional fully nonlinear water waves. Measurements show scalable performance results - in the same order as a dedicated non-library version of the wave tool. Introducing a domain decomposition preconditioner based on a multigrid method, further extends support for multiple GPUs and allows for improvements in performance as well as increased problem sizes.
    Original languageEnglish
    Title of host publication7th International Workshop on Parallel Matrix Algorithms and Applications (PMAA’12) : Programme and Abstracts
    Number of pages0
    PublisherThe European Research Consortium for Informatics and Mathematics
    Publication date2012
    Pages12
    Publication statusPublished - 2012
    Event7th International Workshop on Parallel Matrix Algorithms and Applications (PMAA 2012) - Birkbeck University of London, London, United Kingdom
    Duration: 28 Jun 201230 Jun 2012
    http://www.dcs.bbk.ac.uk/pmaa2012/

    Conference

    Conference7th International Workshop on Parallel Matrix Algorithms and Applications (PMAA 2012)
    LocationBirkbeck University of London
    Country/TerritoryUnited Kingdom
    CityLondon
    Period28/06/201230/06/2012
    Internet address

    Fingerprint

    Dive into the research topics of 'A generic library for large scale solution of PDEs on modern heterogeneous architectures'. Together they form a unique fingerprint.

    Cite this