Manuel Prinz wrote:
Am Donnerstag, den 08.04.2010, 18:22 -0500 schrieb Pavan Balaji:We were having a discussion about a common ABI in the context of MPI-3. But that working group fizzled out. I'll check on what exactly happened with it in the next meeting (in May). Even if the MPI standard doesn't specify ABI compatibility, I'll talk to the Open MPI folks to see if MPICH2 and Open MPI do whatever the ABI working group was proposing outside of the standard, which might be sufficient for you guys.That would be very cool indeed! I had some chat with Jeff Squyres (Open MPI upstream) about that a while ago. Maybe you could put your heads together on this. Having it in the standard would of course be super cool, as it would allow to use other MPI implementations as well without much hassle.[1]
I'll talk to Jeff. This had come up in another meeting last November and Jeff and I talked about doing this. But we never followed up on it; the problem is that this will require quite a few changes in one or both MPI implementations. More importantly, both MPICH2 and Open MPI have derivative implementations (Intel MPI, IBM MPI, Cray MPI, Microsoft MPI, MVAPICH2, etc., for MPICH2, and Sun MPI (renamed to Oracle MPI) for Open MPI) that we would need to coordinate with. This can potentially be a big change for them.
Also, there is some concern amongst MPI developers that we might need to break ABI compatibility once in a while. For example, we recently had to increase the value of MPI_MAX_ERROR_STRING, which breaks ABI compatibility, but was required. So, we need some sort of escape hatch to get around such issues.
I'll try to see what can be done, but no promises :-). -- Pavan -- Pavan Balaji http://www.mcs.anl.gov/~balaji