[Mpi3-rma] [EXTERNAL] Re: MPI-3 UNIFIED model clarification
Barrett, Brian W
bwbarre at sandia.gov
Sun Aug 4 18:10:32 CDT 2013
On 8/4/13 5:02 PM, "Pavan Balaji" <balaji at mcs.anl.gov<mailto:balaji at mcs.anl.gov>> wrote:
On 08/04/2013 05:56 PM, Barrett, Brian W wrote:
So, I guess I really don't see a problem with what we have today. As
soon as you require a target side WIN_SYNC, there's no point in using
Unified; the user and the implementation can live quite happily with
SEPARATE. And, I believe, make SEPARATE go fast on platforms that only
need a memory barrier to ensure target side ordering, rather than a
complicated cache protocol.
That's not true. If two processes are accessing nonoverlapping memory
locations (PUT/GETs and load/stores), it would still be very useful.
With SEPARATE additional synchronization would be required for such
UNIFIED is too complex if the user mostly cares about nonoverlapping
memory regions, and doesn't mind paying a small penalty for overlapping
SEPARATE is too expensive for such cases.
I think we have to agree to disagree on that one, in that I really don't think your example is a it's a common usage model worth destroying UNIFIED over. Most one-sided codes I've seen use one-sided ops for everything in the window, but I haven't seen as many codes as I might like.
Brian W. Barrett
Scalable System Software Group
Sandia National Laboratories
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the mpiwg-rma