source: anuga_core/source/pypar-numeric/README @ 6226

------------------------------------------------
PyPAR - Parallel Python, efficient and scalable 
parallelism using the message passing interface (MPI).

Author:  Ole Nielsen (2001, 2002, 2003)
Email:   Ole.Nielsen@anu.edu.au
Version: See pypar.__version__
Date:    See pypar.__date__

Major contributions by 
    Gian Paolo Ciceri (gp.ciceri@acm.org)
    Prabhu Ramachandran (prabhu@aero.iitm.ernet.in) 
  
Minor but important contributions by
    Doug Orr (doug@arbor.net)    
    Michal Kaukic (mike@frcatel.fri.utc.sk)         
    
------------------------------------------------

The python module pypar.py and the C-extension mpi.c 
implements scalable parallelism on distributed and shared 
memory architectures using the essential subset of 
the Message Passing Interface (MPI) standard.

FEATURES

- Python interpreter is not modified: 
  Parallel python programs need only import the pypar module.

- Easy installation: This is essentially about compiling and linking 
  the C-extension with the local MPI installation. A distutils setup file
  is included.

- Flexibility: Pypar allows communication of general Python objects 
  of any type.  
  
- Intuitive API: 
  The user need only specify what to send and to which processor.
  Pypar takes care of details about 
  data types and MPI specifics such as tags, communicators and buffers.
  Receiving is analogous.
 
- Efficiency:
  Full bandwidth of C-MPI programs is achieved for consecutive Numerical 
  arrays. Latency is less than twice that of pure C-MPI programs. 
  Test programs to verify this are included (pytiming, ctiming.c)
  
- Lightweight:
  Pypar consists of just two files: mpiext.c and pypar.py           

See the DOC file for instructions on how to program with pypar.

    
  
PRE-REQUISITES (on all nodes)

  Python 2.0 or later
  Numeric Python (incl RandomArray) matching the Python installation
  Native MPI C library
  Native C compiler
  
  Pypar has been tested on the following platforms
    MPI on DEC Alpha
    LAM/MPI (6.5.6, 7.0) on Linux (Debian Woody/Cid and Red Hat 7.1) 
    MPICH and LAM 7.0 on Solaris (Sun Enterprise)
    MPICH on Linux
    MPICH on Windows (NT/2000)
    MPICH for Darwin (Mac OS)
    LAM/MPI with Linux on Opteron 64 bit
   
  
INSTALL

  UNIX PLATFORMS        
    Type
      python setup.py install 
 
    This should install pypar and its extension in the default
    site-packages directory on your system.
    
    If you wish to install it in your home directory use
  
      python setup.py install --prefix=~         
      
  WINDOWS (NT/2000, MPICH)
    To build, you need:
      1) MPICH (http://www-unix.mcs.anl.gov/mpi/mpich/)
      2) MinGW (http://www.mingw.org). Tested with GCC v 2.95-3.6.
      3) Set MPICH_DIR to the appropriate directory
      4) Build using 'python setup.py build --compiler=mingw32'
      5) Install using 'python setup.py install'.

  ---
  If you encountered any installation problems, read on.
  Do not worry about messages about unresolved symbols. This is normal 
  for shared libraries on some machines.
    
  To compile C-extension mpi.c requires either a mpi-aware c compiler
  such as mpicc or a standard c compiler with mpi libraries linked in.
  See local MPI installation for details, possibly edit Makefile and 
  type make.   
  See installation notes below about compiling under mpi. 

  
TESTING
  Pypar comes with a number of tests and demos available in the 
  examples directory - please try these to verify the installation.
  

RUNNING PYTHON MPI JOBS 
 
  Pypar runs in exactly the same way as MPI programs written in 
  C or Fortran. 
  
  E.g. to run the enclosed demo script (demo.py) on 4 processors, 
  enter a command similar to
   
    mpirun -np 4 python demo.py
 
  Consult your MPI distribution for exact syntax of mpirun. 
  Sometimes it is called prun and often parallel jobs will be
  submitted in batch mode.
  
  You can also execute demo.py as a stand-alone executable python script 
  
    mpirun -np 4 demo.py
  
  
  Enclosed is a script to estimate the communication speed of your system.
  Execute  
  
    mpirun -np 4 pytiming
  
  Care has been taken in pypar to achieve the same bandwidth and almost as 
  good communication latency as corresponding C/MPI programs.
  Please compile and run ctiming.c to see the reference values:
  
    make ctiming
    mpirun -np 4 ctiming       
    
  Note that timings might fluctuate from run to run due to variable 
  system load.   

  An example of a master-slave program using pypar is available in demo2.py.  
  
  
INSTALLATION NOTES (If all else fails)

Most MPI implementations provide a script or an executable called 
"mpicc" which compiles C programs using MPI and does 
not require any explicitly mentioned libraries. 
If such a script exists, but with a different name, change
the name in the beginning of compile.py. If no such script exists, put
the name of your C compiler in that place and add all required linking
options yourself.

For example, on an Alpha server it would look something like

 cc -c mpi.c -I/opt/Python-2.1/include/python2.1/
 cc -shared mpi.o -o mpi.so -lmpi -lelan
 
 or using the wrapper
 
 mpicc -c mpi.c -I/opt/Python-2.1/include/python2.1/
 mpicc -shared mpi.o -o mpi.so

On Linux (using LAM-MPI) it is 

 mpicc -c mpi.c -I/opt/Python-2.1/include/python2.1/
 mpicc -shared mpi.o -o mpi.so

 Start processors using
 lamboot -v lamhosts  
 
The script install.py (which is using compile.py) is a non-distutils 
way of compiling the extension.


HOW DOES PYPAR WORK?

Pypar works as follows
    1  mpirun starts P python processes as specified by mpirun -np P
    2  each python process imports pypar which in turn imports a shared
       library (python extension mpiext.so) that has been statically linked 
       to the C MPI libraries using e.g. mpicc (or cc -lmpi)
    3  The Python extension proceeds to call MPI_init with any commandline
       parameters passed in by mpirun (As far as I remember MPICH uses this 
       to identify rank and size, whereas LAM doesn't)
    4  The supported MPI calls are made available to Python through the
       pypar interface

    If  pypar is invoked sequentially it supports a rudimentary interface
    with e.g. rank() == 0 and size() == 1
 
 
 
DOCUMENTATION

  See the file DOC for an introduction to pypar.
  See also examples demo.py and pytiming
  as well as documentation in pypar.py. 

  
HISTORY
version 1.9.3 (24 April 2007)
  added bypass form for reduce 
version 1.9.2
  First version under Subversion
version 1.9.1 (15 Dec 2003)
  Consistent naming (all lower case, no abbreviations)
version 1.9 (3 Dec 2003)
  Obsoleted raw forms of communication as they were confusing.
  The semantics of send, receive, scatter, gather etc is now that 
  one can optionally specify a buffer to use if desired.
  Bypass forms added.
  
version 1.8.2 (16 November 2003)
  Fixed scatter, gather and reduce calls to automatically
  figure out buffer lengths. Also tested scatter and gather for
  complex and multidimensional data.
version 1.8.1 (13 November 2003)
  Added direct support (i.e. not using the vanilla protocol) 
  for multidimensional Numeric arrays of type Complex. 
version 1.8 (11 November 2003)
  Changed status object to be (optionally) returned from receive and
  raw_receive rather maintaining a global state where one instance of
  the status object is modified.
  This was suggested by the audience at presentation at Department of
  Computer Science, Australian National University, 18 June 2003.
  
version 1.7 (7 November 2003)
  Added support for Numerical arrays of arbitrary dimension 
  in send, receive and bcast.
  The need for this obvious functionality was pointed by Michal Kaukic.
  
version 1.6.5 (30 August 2003)
  Added NULL commandline argument as required by LAM 7.0 and
  identified by Doug Orr and Dave Reed.
version 1.6.4 (10 Jan 2003)
  Comments and review of installation
version 1.6.3 (10 Jan 2003)
  Minor issues and clean-up
version 1.6.2 (29 Oct 2002)
  Added Windows platform to installation as contributed by 
  Simon Frost
version 1.6.0 (18 Oct 2002)
  Changed installation to distutils as contributed by
  Prabhu Ramachandran
  
version 1.5 (30 April 2002)
  Got pypar to work with MPICH/Linux and cleaned up initialisation
version 1.4 (4 March 2002)
  Fixed-up and ran testpypar on 22 processors on Sun
version 1.3 (21 February 2002)                                      
  Added gather and reduce fixed up testpypar.py
  
version 1.2.2, 1.2.3 (17 February 2002)                                      
  Minor fixes in distribution
version 1.2.1 (16 February 2002)                                      
  Status block, MPI_ANY_TAG, MPI_ANY_SOURCE exported                 
Version 1.2 (15 February 2002)                                             
  Scatter added by Gian Paolo Ciceri                                   
Version 1.1 (14 February 2002)                                             
  Bcast added by Gian Paolo Ciceri                                   
Version 1.0.2 (10 February 2002)                                          
  Modified by Gian Paulo Ciceri to allow pypar run under Python 2.2  
Version 1.0.1 (8 February 2002)                                       
  Modified to install on SUN enterprise systems under Mpich          
Version 1.0 (7 February 2002)                                         
  First public release for Python 2.1 (OMN)                          
  
TODO  
  max_tag is set to 32767. This works for Linux/LAM      
  I couldn't use MPI_TAG_UB as it returned 0.
  I would like a general solution.


  Scatter needs to send buffer specified on all processes
  even though it is ignored by all non-root processes.
  How could we overcome that?
  Similar problem for gather
  
  Gather an scatter: One should be able to specify along which axis
  arrays should be concatenated.


KNOWN BUGS
  Scatter for the moment works only properly when the amount of data is a 
  multiple of the number of processors (as does the underlying MPI_Scatter).
  I am working on a more general scatter (and gather) which will
  distribute data as evenly as possible for all amounts of data.  

      
LICENSE
    This program is free software; you can redistribute it and/or modify
    it under the terms of the GNU General Public License as published by
    the Free Software Foundation; either version 2 of the License, or
    (at your option) any later version.

    This program is distributed in the hope that it will be useful,
    but WITHOUT ANY WARRANTY; without even the implied warranty of
    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
    GNU General Public License (http://www.gnu.org/copyleft/gpl.html)
    for more details.

    You should have received a copy of the GNU General Public License
    along with this program; if not, write to the Free Software
    Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307


 Contact address: Ole.Nielsen@anu.edu.au

 =============================================================================
   
ACKNOWLEDGEMENTS  
  This work was supported by School of Mathematical Sciences at 
  the Australian National University and funded by Australian 
  Partnership for Advanced Computing (APAC).
  Many thanks go to 
  
    Gian Paolo Ciceri (gp.ciceri@acm.org) 
        for fixing pypar to run under Python 2.2 and for adding all the 
        collective communication stuff from version 1.1 and onwards.
    Prabhu Ramachandran (prabhu@aero.iitm.ernet.in)
        for making a proper distutils installation procedure
    Simon D. W. Frost (sdfrost@ucsd.edu)
        for testing pypar under Windows and adding installation procedure 
    Jakob Schiotz (schiotz@fysik.dtu.dk) and Steven Farcy (steven@artabel.net) 
        for pointing out initial problems with pypar and MPICH.
    Markus Hegland (Markus.Hegland@anu.edu.au)
        for supporting the work.
    Doug Orr (doug@arbor.net)
        for identifying and fixing problem with LAM 7.0 requiring 
        last commandline argumnet to be NULL.
    Dave Reed (dreed@capital.edu)
        for pointing out problem with LAM 7.0 and for testing the fix.  
    Matthew Wakefield (matthew.wakefield@anu.edu.au)
        for pointing out and fixing support for Mac OS X (darwin)
    David Brown (David.L.Brown@kla-tencor.com)
        for contributing to the Windows/Cygwin installation procedure   
    Michal Kaukic (mike@frcatel.fri.utc.sk)     
        for pointing out the need for multi dimensional arrays and 
        for suggesting an implementation in the 2D case.