------------------------------------------------
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.
Download Latest Version
pypar-2.1.4_94.tgz (114.8 kB)
