Hey now, blocking doesn't mean that MPI /must/ block, it only means that it /can/ block. Blocking and synchronizing are certainly not the same; Ssend synchronizes, which is a stronger semantic than blocking. ~Jim. On Wed, Jun 12, 2013 at 5:07 PM, Jeff Hammond <[email protected]> wrote:
Sure, let's separate these issues. I think Pirate RMA (RRPUT, RRACCUMULATE and RRRGET_ACCUMULATE) is a more _local_ solution in the MPI standard sense :-)
SEND is also a nonblocking operation because the implementation can always buffer the data and return before the remote RECV is posted. Only SSEND is blocking if one asserts that blocking is equivalent to "requires remote agency to complete".
So basically, I completely agree with the Forum back-pushers that IWIN_FLUSH etc. are as useful as ISEND.
Jeff
On Wed, Jun 12, 2013 at 4:58 PM, Pavan Balaji <[email protected]> wrote:
The push-back at the Forum was that WIN_FLUSH and WIN_UNLOCK are local operations (in MPI's definition of local), so they are already
nonblocking.
I was not at the Forum. I pointed out to Wesley after he returned that they are only local in the simple case where no one else is holding the lock. Otherwise they are not.
Further, the proposal is not only about WIN_FLUSH and WIN_UNLOCK. It was for a long list of blocking functions: COMM_ACCEPT, COMM_CONNECT, WIN_FENCE, WIN_COMPLETE, WIN_WAIT, WIN_UNLOCK, WIN_FLUSH, WIN_FLUSH_LOCAL, WIN_FLUSH_ALL, WIN_FLUSH_LOCAL_ALL, ... The point was that MPI should deal with concepts, not specific function exceptions. Nonblocking communication is one such concept, and it should be valid for all operations (that are not local).
FWIW, my proposal was only to add nonblocking variants to blocking calls. Adding extra per-op remote completion requests is a separate issue and should not be mixed with this. I personally think it's useful (in fact, my original proposal for RPUT/RGET equivalents had it), but that's still a separate issue.
-- Pavan
On 06/12/2013 09:41 AM, Jeff Hammond wrote:
So Wes was telling me about IWIN_FLUSH yesterday (this function is a great idea, btw) and the alternative of adding double-request RPUT and RACCUMULATE that provide a request for both local and remote completion. By induction, this would imply a triple-request RGET_ACCUMULATE operation, which I also think is a good idea. It's really silly to have to wait for the result buffer to be ready to reuse the origin buffer in RGET_ACCUMULATE. It seems that many implementations will be able to send the entire origin buffer a nontrivial amount of time before the result comes back, particularly if the remote side doesn't have hardware-based progress and the data has to wait in a network buffer.
I'm sorry that I missed the last day of the Forum in San Jose. I didn't know these operations were on the table. They sound like a great idea to me.
Jeff
-- Pavan Balaji http://www.mcs.anl.gov/~balaji
-- Jeff Hammond Argonne Leadership Computing Facility University of Chicago Computation Institute [email protected] / (630) 252-5381 http://www.linkedin.com/in/jeffhammond https://wiki.alcf.anl.gov/parts/index.php/User:Jhammond ALCF docs: http://www.alcf.anl.gov/user-guides _______________________________________________ mpi3-rma mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma