MPI implementations

[Konrad Hinsen]

On Mar 18, 2006, at 5:04, Daniel Nogradi wrote:

> I plan to use python parallel and -- fortunately or unfortunately --
> there are several implementations. At the moment I'm aware of
>
> http://datamining.anu.edu.au/~ole/pypar/
> http://pympi.sourceforge.net/
> http://www.fysik.dtu.dk/~schiotz/comp/PythonAndSwig/pythonMPI.html
>
> and wondered if anyone had experience with any of them. Pros, cons?
> Did anyone try more than one?

Add this one:

	http://dirac.cnrs-orleans.fr/ScientificPython/ScientificPythonManual/ 
Scientific_27.html

and unless you are 100% convinced that MPI is the interface you want  
to use, also check out this:

	http://dirac.cnrs-orleans.fr/ScientificPython/BSP_Tutorial.pdf

PyPar and pyMPI are both high-level interfaces to MPI, i.e. they  
propose an MPI-like message passing interface with the main added  
value being communication of arbitrary Python objects. My own MPI  
interface in ScientificPython is more low-level, since it handles  
only strings and arrays. Its strong point is its C interface, which  
makes it possible to write parallelized mixed C-Python code and to  
integrate parallelized C libraries into Python projects.

My BSP interface (the second link) is what I personally consider the  
most convenient interface for parallel programming in Python. It is  
built on the BSP (Bulk Synchronous Processing) model that works at a  
higher level than message passing: the programmer does not have to  
worry about synchronization, which is a major source of hard-to-track- 
down bugs.

Konrad.
--
---------------------------------------------------------------------
Konrad Hinsen
Laboratoire Léon Brillouin, CEA Saclay,
91191 Gif-sur-Yvette Cedex, France
Tel.: +33-1 69 08 79 25
Fax: +33-1 69 08 82 61
E-Mail: konrad.hinsen at cea.fr
---------------------------------------------------------------------