So then, where would you draw the line on the old mpi datatype
access issue?
instantiation" facility is certainly a 3rd party, added-value product.
This is a change in thinking for me, but I admit it makes more sense
in the long run.
To implement this capability I need to have an "on the wire"
description of a compound datatype. Certainly the program which
generated the datatype is an expert in describing the internals of the
type. For the description (string) to be portable a there needs to be
an agreed upon standard specification. Therefore, I identify the
specification of a datatype description string as a role for the
Forum.
Whether or not MPI2 needs to additionally provide an MPI_Get_char_datatype()
function to generate one of these descriptions is an open question.
It seems to me that a 3rd party would have to implement MPI in its entirety
just to create this function; i.e., this function needs access to
"hidden" MPI implementation internals. But perhaps there's another way?
If not, I also regard it as a target for MPI2. :-(
Richard Frost
SDSC
On Thu, 8 Feb 1996, Richard Frost wrote:
>
> In my opinion, MPI-IO should view itself as an added-value effort to MPI --
> a package containing a library and utility programs for developers. As such,
> the MPI-IO needs a "standard" specification. Implementation is great,
> but getting it right is even more important.
>
> MPI offers a basis for a new generation of software libraries.
> Trying to add all of the world's computing concerns to the MPI standard
> would be a mistake -- implementation would be too great a chore.
> Those developing add-on products might (and have) discover some critical base
> functionality which the Forum overlooked. Making the distinction between
> what is critical for the base specification is the role of the MPI Forum.
>
> Those of you who are directly involved with MPI-IO implementations
> are driven by real, pressing needs in various development efforts.
> However, in the process of implementation some individuals have attempted
> to add functionalities which are highly desirable in message-passing programs
> but orthogonal to an I/O specification.
>
> I urge such individuals to contact the MPI Forum and learn who in the
> community is working on similar efforts. For example, a generalized
> distributed tensor data library based on MPI would be very useful
> but should be a separate added-value effort. Hopefully, those who have
> expressed an interest in that area will work closely with the community
> to get that specification right too.
>
> See you at the March forum meeting.
>