Question

It’s all begun when I needed to MPI_Bcast a 64 bit integer. Since MPI does not know how to handle it I did:

template<typename T>
inline int BcastObjects(T* pointer, 
                        int count, 
                        int root, 
                        MPI_Comm comm)
{
    return MPI_Bcast(pointer, 
                     count * sizeof(*pointer), 
                     MPI_BYTE, 
                     root,
                     comm);
}

Now I can do:

int64_t i = 0;
BcastObjects(&i, 1, root_rank, some_communicator);

Then I started to use BcastObjects to send over an array of structures. I wonder if it’s OK to do that?

The manuals about MPI_Datatype focus on how to do it, but not on why would I want to do it.

Answer 1

Why not just use MPI_INT64_T?

You can always mock up your own datatypes with MPI_Byte or what have you; the datatype stuff is there so that you don’t have to. And in many cases it’s much easier; if you want to send data that has “holes” in it (eg, a slice of a multidimensional array, data in a structure that has gaps), you can map that out fairly straighforwardly with a datatype, whereas you’d have to manually count out byte strings and use something like MPI_Pack otherwise. And of course describing the data at a higher level is certainly less brittle if something in your data structure changes.