MPI_Intercomm_create functionality

[rabauke]

MPI_Intercomm_create functionality

[rabauke]

@a-jp: MPI inter-communicators are currently not supported. However, this is certainly a feature worth considering.

Although MPI handles all communicators via the same kind of handle, MPL should model inter-communicators by a specific dedicated class, not via mpl::communicator. Thus, this will not be implemented just via a minimal modification of the mpl::communicator class.

One reason, why inter-communicators are not supported by MPL is just that I never had an actual use-case for this. I know the examples from the MPL standard, which are, however, rather academic. Can you please explain your use-case?

[a-jp]

@rabauke many thanks for the comments. The use-case here is multiple program, multiple data (MPMD) style launch using: mpirun -np X prog_1 : -np Y prog_2.

In this case comm_world is of size X+Y. Within comm_world two communicators are created:

mpl::communicator group0(mpl::communicator::split, comm_world, 0);

Here 0 is the application number, for prog_1, and group0 is of size X. A second communicator is created as:

mpl::communicator group1(mpl::communicator::split, comm_world, 1);

Here 1 is the application number, for prog_2, and group1 is of size Y.

The use-case is coupling two codes together that can be used to solve a complex multi-domain problem, that is otherwise not solvable via either individual code. The coupling between the codes is achieved via MPI. In this use-case it is essential to be able to communicate between group0 and group1, hence the need for MPI_Intercomm_create.

An older paper, as this is a well established approach, and very relevant reference for this use-case is below [1]. This represents a real use-case and is representative of the type of problem that interests me; namely coupling two codes via MPI.

I hope you consider this for inclusion into your library, for present I will write a simple wrapper class around MPI_Intercomm_create type functions as I go.

Many thanks,

[1]

[rabauke]

@a-jp Thanks for further details. I will consider inter-communicators as a future feature of MPL.

[a-jp]

ok thanks

[VictorEijkhout]

On , 2021Oct3, at 04:26, a-jp ***@***.******@***.***>> wrote: In this use-case it is essential to be able to communicate between group0 and group1, hence the need for MPI_Intercomm_create. I understand what you’re doing. But because your group0 & 1 come from splitting com_world you still have comm_world as a means of communication. Why is the inter communicator a better mechanism? Other usage scenario: an inter communicator is created from MPI_Comm_spawn. I have a user who does this spawning but I can’t give you the details. However, I must say that this is a very rare case. Victor.

[rabauke]

This should stay open until milestone has been reached.