New MPI Send

Anna Rounbehler - Technical Staff (anna@sky.com)
Mon, 17 Jun 96 17:08:02 EDT

I took some time to think about some of your suggestions, which I thought
may have captured issues that many of us have been mulling over.
Your ideas have been very helpful, since they cause us to think about
the balance between generic and specific.

One reason why MPI-2 extensions are valubale is that they provide
a vehicle for unique and specific MPI functionality. If the basic
MPI functions grow too large in scope, some implementors end up with
too much overlap between extensions and MPI-1 functionality. A new MPI_
Xsend command may fall into this category. Some shared memory architectures
encapsulate RMA, DMA, and local transfers as part of an implementation
layer for the standard MPI send commands. An additional MPI Send
encourages a decoupling of data transfer type, at a high level. It may be
more implementation & user friendly to encapsulate this functionality in
the MPI-2 extensions as unique function calls, instead of an MPI basic
function.

There is a fine line between an advantage derived from adding MPI basic
functions/extensions and leaving room for the implementators. I suspect
this line wiggles around due to the variety of architectures and users
that MPI supports.

If sometimes a proposal is rejected, it is not because it is poor, but
possibly because the advantage to some is an equal disadvantage to others.
That makes it a tough call. So I think, therein lies the balance.