I believe that InfiniBand would have to copy the entire buffer before starting the operation. It does not seem to be beneficial to bury that kind of performance anomaly. It would be sort of like having MPI_Sendrecv with an "in place" option. The best a network could ever do is to break it up into some "safe" size for the network and do it in pieces with round trips after each piece completed. Keith
-----Original Message----- From: Jeff Hammond [mailto:[email protected]] Sent: Thursday, March 31, 2011 8:25 PM To: MPI 3.0 Remote Memory Access working group Cc: Underwood, Keith D Subject: Re: [Mpi3-rma] GetAccumulate restriction needed?
What are the use cases that make implementing this difficult? I can argue both sides of this, so I'm curious what you think is the deal-breaker.
Thanks,
Jeff
On Thu, Mar 31, 2011 at 9:17 PM, Underwood, Keith D <[email protected]> wrote:
Hi All,
Brian and I were having a discussion about something else today and realized that we weren't sure of something very important:� is there anywhere in the one-sided chapter that prohibits the source buffer of the GetAccumulate and the reply buffer of the GetAccumulate from being the same buffer?� It seems like this should be prohibited.
Keith
_______________________________________________ mpi3-rma mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma
-- Jeff Hammond Argonne Leadership Computing Facility [email protected] / (630) 252-5381 http://www.linkedin.com/in/jeffhammond