Techniques used to implement a
dataparallel matrix class library (SymPLA) in C++,
for a sequential frontend
working with a parallel backend

Thesis for the degree of Cand. Scient.
by
Yngve Nysæter Pettersen
=2.5cmpsfigs/uib.ps
Department of Informatics
University of Bergen
Norway

• Contents
• List of Figures
• Introduction
  • Background
  • An overview of SymPLA
• Parallel programming tools
  • Parallel programming languages
    • High Performance Fortran
    • MasPar Programming Language
    • Parallel C++
      • pC++
      • Compositional C++
  • Message passing environments
    • PVM
    • MPI
    • NX
    • BLACS
  • Numerical libraries
    • BLAS
    • LAPACK
    • SCALAPACK
• The SymPLA class library
  • Properties
  • Basic foundations
  • Design and implementation
    • Matrix declaration
    • Matrix algebra
    • Accessing the matrix size
    • Maximal or minimal elements
    • How to store the matrix data
      • Modifications in other structures
    • Submatrices
      • IndexLists
      • Practical submatrices
    • Factorization and system solving
      • Implementation details
    • Input and output of data in matrices
      • Data in files
      • Data in memory
      • Constants
      • The Identity matrix
      • The diagonal matrix
      • Element functions
  • Vectors
  • Summary of techniques
• Implementations
  • Sequential version
  • Parallel versions
    • SIMD implementation
    • MIMD implementation
• Experiences
  • Performance
    • LAPACK version
      • Addition
      • Multiplication
      • LU factorization
      • Overhead
    • ScaLAPACK version
      • Addition
      • Multiplication
      • LU factorization
      • Overhead
      • Scalability
    • MasPar version
      • Addition
      • Multiplication
      • LU factorization
      • Overhead
  • Ease of use
    • A conjugated gradient routine
  • Portability and extensions
• Conclusions and future work
  • A complete, standard Matrix class library
  • Parallel extensions
  • Conclusions
• References
• Source code: SymPLA header files
• Source code: SymPLA Library
• Source code: LAPACK version
• Source code: SCALAPACK version
• Source code: MasPar version
• About this document ...

Forskningsdokumentasjon ved Universitetsbiblioteket i Bergen