Introduction

• Two standard programming models for parallelisation:
  message passing:
  
  basic idea: exchange of messages between independent tasks

  • + well established standards
  • + high portability (from PC cluster to supercomputer)
  • + implicit synchronisation with messages
  • - complete redesign necessary
  • - lots of nasty details ("parallel assembler")
  shared memory:
  
  basic idea: loop parallelisation with compiler directives

  • + incremental approach possible, only parallelise hot spots
  • + supported by compiler
  • + programming model similar to scalar programs
  • - code not portable
  • - explicit synchronisation necessary
• OpenMP:
  industrie standard describing directives for shared memory parallelisation
  controlled by OpenMP Architecture Review Board
  members:

  • hardware providers (Compaq/Digital, HP, Intel, IBM, Sun, ...)
  • software providers (KAI, NAG, ANSYS, Genias, ...)
  standard documents:

  • Fortran (Oktober 1997)
  • C/C++ (Oktober 1998)
  further development:

  • Fortran additions (array syntax)
  • parallel I/O
• Web pointer:
  http://www.openmp.org/
  homepage of OpenMP organisation
  Fortran , C/C++
  local copies of standard documents (in PDF)
  Fortran , C/C++
  user manuals for OpenMP compiler from KAI
  all tutorial example programms
  http://www.openmp.org/presentations/index.cgi?sc98_tutorial
  http://www.npac.syr.edu/users/gcf/dbppindod/dbpp2openmp/index.htm
  http://www.psc.edu/training/PPT_Oct_98/openmp/openmp1.htm
  tutorials