[Numpy-discussion] Best way to run python parallel

Mandus mandus@gmail....
Fri Mar 30 05:59:20 CDT 2007


We have also done some work with BSP in the past (actually together
with Konrad). It's a great model, and quite comfortable to work with. Also,
with Konrads implementations it was very efficient at sending Numeric
arrays around.

But the main problem with BSP is that it is not very much used in the
community, as Brian points out. This mean that if you get access to
some parallel computer, it is highly unlikely that BSP is available to you.
We also had a hard time building the stuff on our own computers, as the
libraries was already quite old back when we did this stuff.

About the BSP model itself, as Brian points out, its usablility may depend
on the kind of problems you need to solve. If you need to solve PDEs
(which is what we did with it) and these PDEs gives sparse linear systems
using for instance finite element mehthods and linearization - then BSP can
be quite efficient. For some problems, we actually found that python & BSP
performed on par with C & MPI, both in terms of total runtime and parallel
efficiency. No doubt, that's not a general conclusion, but it certainly
shows the possibilites.



On 3/30/07, Brian Granger <ellisonbg.net@gmail.com> wrote:
>
> We looked at the BSP model at various points in implementing the
> parallel IPython stuff.  While I wouldn't say that IPython uses a BSP
> model, there are some similarities.  But in the broader realm of
> scientific computing, BSP has never really caught on like MPI has - in
> spite of having some nice ideas like being able to predict the
> performance of a parallel BSP code.
>
> The main difficulty is that BSP is a much more limited model that MPI
> - which is why you can predict the performance of BSP using codes.
> The main limitation is that communication and computation cannot
> overlap.  For some codes that is fine and there might be benefits for
> using BSP over MPI.  For general parallel codes and algorithms
> however, you do want to overlap communications and computation.  In
> fact, if you don't you can severely limit the scalability of your
> code.
>
> Brian
>
>
>
>
>
> On 3/29/07, Sebastian Haase <haase@msg.ucsf.edu> wrote:
> > Hi,
> > What is the general feeling towards  BSP on this list !?
> > I remeber Konrad Hinsen advertising it on the SciPy workshop '03 .
> > It is supposed to be much simpler to use than MPI, yet still powerful
> > and flexible enough for most all applications.
> > It is part of Konrad's ScientificPython ( != SciPy )
> >
> > Some links are here:
> > http://www.bsp-worldwide.org/
> > http://en.wikipedia.org/wiki/Bulk_Synchronous_Parallel
> >
> > Evaluating Scientific Python/BSP on selected parallel computers
> > http://ove.nipen.no/diplom/
> >
> > http://dirac.cnrs-orleans.fr/plone/software/scientificpython/
> >
> > - Sebastian Haase
> >
> >
> >
> > On 3/29/07, Peter Skomoroch <peter.skomoroch@gmail.com> wrote:
> > >
> > >
> > > If you want to use PyMPI or PyPar, I'm writing a series of tutorials
> on how to
> > > get them running on Amazon EC2,
> > >
> > >
> http://www.datawrangling.com/on-demand-mpi-cluster-with-python-and-ec2-part-1-of-3.html
> > >
> > >
> > > I'm using PyMPI on a 20 node EC2 cluster and everything seems groovy,
> but I'm
> > > relatively new to MPI, so I have probably overlooked some easier
> solutions.
> > >
> > > Any feedback on the writeups from Python gurus would be appreciated.
> > >
> > > -Pete
> > _______________________________________________
> > Numpy-discussion mailing list
> > Numpy-discussion@scipy.org
> > http://projects.scipy.org/mailman/listinfo/numpy-discussion
> >
> _______________________________________________
> Numpy-discussion mailing list
> Numpy-discussion@scipy.org
> http://projects.scipy.org/mailman/listinfo/numpy-discussion
>



-- 
Mandus
The only dr. Mandus around.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://projects.scipy.org/pipermail/numpy-discussion/attachments/20070330/f94ad0d5/attachment-0001.html 


More information about the Numpy-discussion mailing list