Open MPI logo

Open MPI User's Mailing List Archives

  |   Home   |   Support   |   FAQ   |   all Open MPI User's mailing list

From: Toon Knapen (toon.knapen_at_[hidden])
Date: 2005-03-10 09:58:07


As posted on comp.parallel.mpi, I also wanted to forward this message to
users_at_[hidden] because I think it is relavent to the (undoubtly
upcoming) mpich2 <-> open-mpi discussion.

<quote>
Greg Lindahl wrote:
> The first question is: Does an ABI provide enough benefit for people
> to care?

I care a _lot_ !

> To care enough to sit on a committee?

Abolutely. Only we're a small SME and thus we need to strategically
choose the 'fights' that are most important to us. I would really want
to join but I'm afraid that a small SME as ours can do little against
the giants that are involved in MPI.

I sure hope you go ahead with this idea and will certainly follow the
discussion.

OTOH I once heard Intel is working on defining a ABI in cooperation with
the mpich guys. I believe mpich2 also should allow to change the backend
(specific to the switch) when launching the application. OpenMPI also is
working on this but OpenMPI and mpich2 will not be compatible however.
So there are several efforts but now I think it's important to sync them.

Greg Lindahl wrote:
> Recently there's been a little chatter on the Beowulf cluster computing
> mailing list about an MPI ABI, an application *binary* interface, so that
> programs could automagically be used with any MPI chosen at run-time,
> instead of today's compile-time choice.
>
> The first question is: Does an ABI provide enough benefit for people
> to care? To care enough to sit on a committee?
>

> If the answer is "yes", then I think we'll have one. The minimum
> technical issues revolve around the contents of <mpi.h> and the names
> of shared libraries. The amount of work for MPICH or OpenMPI to
> support that part of an ABI is modest.
>

> If we wanted to go farther, I have a strawman proposal which addresses
> a generic startup procedure which would allow user applications, MPI
> implementations, and queue systems to all live in peace and harmony.
>

> This talk, which I gave last week at the OpenIB meeting:
>

> http://www.openib.org/docs/oib_wkshp_022005/mpi-abi-pathscale-lindahl.pdf
>

> mostly talks about why we need an ABI, who wins and loses as a result
> of having one, and the pieces that could be in it. Please give it a
> look.
>

> -- greg

</quote>