[MPI3 Fortran] MPI Data types
longb at cray.com
Thu May 7 15:54:41 CDT 2009
Jeff Squyres wrote:
> On May 7, 2009, at 1:30 PM, Jeff Squyres wrote:
>> This brings up the struggle I brought before: why does Fortran need
>> the (count, MPI_Datatype) tuple? Because it's part of MPI, that's
>> why -- it's the standard.
> I let out an important piece of rationale here: part of the whole
> reason to have MPI_Datatypes is to allow the MPI implementation to
> optimize the memory layout map *once*, and then just use that
> optimized map forevermore. If you have no MPI_Datatype argument (that
> is a handle to this optimized map), then you have to construct
> datatypes (and optimize them) on the fly. This has shown, in
> real-world MPI codes, to be quite expensive (read: slow).
Thanks for this clarification. Useful information. While you're on a
roll, I have another question:
Is MPI_Datatype intended to describe just one instance of a structure
(i.e. the layout of its internal, maybe not contiguous pieces)? Or
could it be used to describe a collection of such objects, such as a
subset of the elements of an array, where each element is a structure?
The second option is the potentially more interesting one for Fortran users.
Bill Long longb at cray.com
Fortran Technical Support & voice: 651-605-9024
Bioinformatics Software Development fax: 651-605-9142
Cray Inc., 1340 Mendota Heights Rd., Mendota Heights, MN, 55120
More information about the mpiwg-fortran