/dplasma

DPLASMA is the fastest implementation of a dense linear algebra package for distributed heterogeneous systems. It is designed to deliver sustained performance for distributed systems where each node featuring multiple sockets of multicore processors, and if available, accelerators, using the PaRSEC runtime as a backend.

Primary LanguageCOtherNOASSERTION

DPLASMA is the leading implementation of a dense linear algebra package for distributed, accelerated, heterogeneous systems. It is designed to deliver sustained performance for distributed systems where each node featuring multiple sockets of multicore processors, and if available, accelerators like GPUs or Intel Xeon Phi. DPLASMA achieves this objective through the state of the art PaRSEC runtime, porting the Parallel Linear Algebra Software for Multicore Architectures (PLASMA) algorithms to the distributed memory realm.

Table of Contents

Features

DPLASMA’s feature set includes:

  • Linear systems of equations (Cholesky, LU [inc. pivoting, PP], LDL [prototype]).
  • Least squares (QR & LQ).
  • Symmetric eigenvalue problem (Reduction to Band [prototype]).
  • Level 3 Tile BLAS (GEMM, TRSM, TRMM, HEMM/SYMM, HERK/SYRK, HER2K/SYR2k.
  • Covers double real, double complex, single real, and single complex (D, Z, S, C) precisions.
  • Provides ScaLAPACK-compatible interface for matrices in F77 column-major layout
  • Supports: Linux, Windows, Mac OS X, UN*X (depends on MPI, hwloc)

History

DPLASMA started as the distributed version of ICL PLASMA linear algebra package implemented on top of the PaRSEC runtime, a generic framework for architecture aware scheduling and management of micro-tasks on distributed many-core heterogeneous architectures. For a long period, the DPLASMA library was packaged as part of the PaRSEC project. DPLASMA is now an independent project and uses PaRSEC as a dependency.

Installation

Please read the INSTALL.rst file for the software dependencies and the installation instructions. In addition, the PaRSEC wiki might contain additional information.

Links

When referencing DPLASMA in a publication please use the following reference: [10.1109/IPDPS.2011.299]: Flexible Development of Dense Linear Algebra Algorithms on Massively Parallel Architectures with DPLASMA [bibtex].

If you are interested in this project and want to join the users community, our mailman will be happy to accept you on the project user (moderated) mailing list at dplasma-users@icl.utk.edu.

Happy hacking,
The PaRSEC team.