[Mpi3-rma] Fix for win_allocate_shared
htor at illinois.edu
Fri Sep 23 12:39:10 CDT 2011
On Fri, 2011-09-23 at 12:35 -0500, Pavan Balaji wrote:
> On 09/23/2011 12:32 PM, Torsten Hoefler wrote:
> > On Fri, 2011-09-23 at 12:25 -0500, Pavan Balaji wrote:
> >> On 09/23/2011 02:39 AM, Torsten Hoefler wrote:
> >>> "The semantic rules for shared memory windows in the MPI_RMA_SEPARATE
> >>> model allow concurrent load/store accesses from two distinct processes
> >>> to the same target window that is remote to both processes. An
> >>> implementation cannot expose shared memory windows on hardware that does
> >>> not support this access mode."
> >> This seems to suggest that it won't work for non-cache-coherent systems.
> >> Didn't we decide against that?
> > Brian probably wants to comment on this :-).
> > Here's my recap: The issue was that many architectures may not be able
> > to support UNIFIED. So we don't want to limit the SEPARATE model too
> > much in order to still make it useful. If we later, based on our
> > experience, find that we need this limitation to support existing
> > architectures, then we can add the restriction. However, if we add it
> > now, it may limiting the programming model unnecessarily.
> I'm not sure how this is better than my original suggestion of leaving
> this as "undefined" for SEPARATE. That way, it is possible to define it
> in the later MPI standards. But Jim and you didn't want to do that.
It clarifies semantics for separate (the first part) and the advice to
implementers makes the semantic restriction explicit.
More information about the mpiwg-rma