My point was absolutely not to say that it cannot be done one way or another (I actually use Hobbes enclaves and DTK to do this right now). I have been asked to mail this use case to the mailing list since the point was raised during the last MPI forum meeting that MPI sessions could be used for application composition, and I did so.

You can probably do this today with wi4mpi ( https://github.com/cea-hpc/wi4mpi). Jeff On Wed, Aug 9, 2017 at 7:58 AM, Vallee, Geoffroy R. <valleegr at ornl.gov> wrote: > Hi,

...

To follow up on a discussion we had a few weeks ago, here is a brief description of a use case where multiple implementations of MPI could be used by a single process. This is more an illustration of the idea rather than a detailed description of a model.

Let’s assume we are composing a multi-physics applications where each kernel, that has been developed separately, is an MPI application and requires a specific implementation of MPI (e.g., it requires a customized flavor of Trilinos, ultimately creating a dependency to OpenMPI). Let’s also assume that a composition library is based MPI sessions and rely on a specific implementation of MPI. It is then possible to compose the multi-physics application by coupling the two MPI codes with the composition library, the two MPI implementations being potentially different, in which case a single process is, in fact, using two different MPI implementations.

Let me know if you have any question.

Thanks, _______________________________________________ mpiwg-sessions mailing list mpiwg-sessions at lists.mpi-forum.org https://lists.mpi-forum.org/mailman/listinfo/mpiwg-sessions

-- Jeff Hammond jeff.science at gmail.com http://jeffhammond.github.io/