Move functionality to submodules

[simonbyrne]

Should we make submodules match MPI function groupings? I've done this for certain additions (MPI.Types for MPI_Type_XXX functions, MPI.File for MPI_File_XXX functions). Should we do this for others? e.g. MPI_Comm_XXX, MPI_Cart_XXX, MPI_Win_XXX.

This is slightly complicated by the fact that Comm and Win are already used by type names. So in that case we would either need to rename the types (e.g. to Communicator and Window), or use different names for the modules (one convention in Julia is to use the plural name for the module, e.g. Comms, so that the function would be MPI.Comms.rank).

Thoughts?

[lcw]

I am okay with making submodules match MPI function groupings. The main thing I would like to see is that the interface is consistent. So if we do it in one place we do it everywhere, if possible.

[simonbyrne]

I agree. I chose MPI.Types instead of MPI.Type to avoid conflicts with Type{X} (I thought it was a reserved word, but it isn't, so we could work around it by referring to Base.Type{T} where necessary).

[lcw]

At first glance MPI.Comms.rank feels weird but if we are consistent everywhere (i.e., use the plural name for every module) then the interface should be fairly discoverable.

[ViralBShah]

It really should be MPI.Comm.rank. Comms just sounds really wrong.

[PhilipVinc]

Since we're talking about changing the MPI naming convention, and Julia has dispatch:
Why shouldn't MPI.rank(::MPI.Comm.Communicator) just work? Of course calling MPI.rank on an array makes no sense, so MPI.Comm.rank is ultimately redundant...

I mean, MPI is a package by itself. I believe submodules should be used in this context to:
a) avoid namespace pollution (not relevant in this case, I believe)
b) Make the names more akin standard MPI name (we can still have this, by using MPI.Comm.rank inside of MPI )
c) Separate functionality (can be done, see (b))

[simonbyrne]

If we do name the module Comm, what do people prefer for the type name: Communicator (vote 🎉) or CommHandle (vote 🚀)?

[eschnett]

People will expect the type name to be one of MPI_Comm, MPI.Comm, Comm, etc. Changing this is very invasive.

Changing the function name is more easily possible. I think that object-oriented MPI bindings are not a good pattern to follow because Julia has multiple dispatch. In C++, one would write comm.rank() (where comm is an object of type MPI::Comm). The Julia equivalent is rank(comm), or MPI.rank(comm) (depending on how the module is imported).

[ViralBShah]

I would be ok with not doing submodules.

[lcw]

I am okay without submodules as well and like MPI.rank(comm).

[simonbyrne]

I can live with that.