[Mpi3-rma] [EXTERNAL] Re: MPI-3 UNIFIED model clarification

[Barrett, Brian W]

On 8/4/13 5:44 PM, "Pavan Balaji" <balaji at mcs.anl.gov<mailto:balaji at mcs.anl.gov>> wrote:

On 08/04/2013 06:33 PM, Barrett, Brian W wrote:
I'm either parsing your first or second sentence incorrectly.  Are you
saying that you're doing a get in which the initiator-side address is in
a window and that reading the data after it's arrived is a load and
therefore causes problems for separate?  Or are you saying that you
believe that puts/gets which move memory within a window are considered
loads?

I'm saying the first one, yes.  I didn't understand what you mean by the
second one, so I can't say for sure, but I think my answer is yes for
that one as well.

For the second question, if I do a PUT from a local window memory to a
remote window memory, it is a load on the local window memory.  This is
based on the MPI-3 standard page 436 lines 29-30 (though it says
MPI_SEND, I think it is valid for MPI_PUT as well).

Ah, ok.  I did not interpret that paragraph that way.  Since one-sided operations are explicitly not called out as being incoherent, I interpreted it as the one-sided operations must do the "right thing".  But I can see your concern.  Not sure what the right solution is: 1) fix the specification of separate to do the right thing for one-sided operations 2) loosen the unified specification to work in more places, or 3) add a third memory model.  1) has some non-cache coherent issues in my mind.  2) means we lose SHMEM-like semantics on those platforms, which greatly worries me.  3) might be the best option (ugh).

Brian

--
  Brian W. Barrett
  Scalable System Software Group
  Sandia National Laboratories
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mpi-forum.org/pipermail/mpiwg-rma/attachments/20130804/70fb53dc/attachment-0001.html>