[IPython-user] Initializing MPI

Stefan Seefeld seefeld@sympatico...
Thu Nov 12 13:04:24 CST 2009


On 11/12/2009 01:32 PM, Brian Granger wrote:
> Stefan,
> Sorry, but I am horribly busy today.  But here is a short answer.

Thanks for the quick answer !

> * Yes, you can use anything you want to initialize MPI.
> * If you want the engine ids to match the MPI ranks, you will need to use
> an ipython config file to add support for your method of initializing MPI.
> See IPython/kernel/scripts/ipengine.py and IPython/kernel/config/
> for details and examples.
> Note: I have completely refactored how ipcluster works in a branch.  
> It will be
> much easier to config all  of this in the near future.  But the 
> overall idea will still
> be the same.
> Things will be better for me next week, but please don't hesitate to 
> ask quesitons.

OK, I can certainly live with having to use mpi4py in the short term. I 
just want to avoid an additional dependency. (In particular as Fedora 
provides SciPy and IPython packages, but no mpi4py, making it a little 
hard for end users to set up their machine.)

I'm still wondering what options I need to pass to MPI_Init(). I tried 
running MPI_Init, passing the content of sys.argv through. But that had 
the effect that all of my module instances reported the same rank '0', 
which seems to imply they all had no knowledge of each other. I figure 
to tell them about the other processes, and set up the channels, 
requires some argument to MPI_Init(), which would normally be provided 
by mpirun, but which I can't find any traces of in sys.argv.



       ...ich hab' noch einen Koffer in Berlin...

More information about the IPython-user mailing list