[mpiwg-rma] summary of ticket #398 discussion
Jeff Hammond
jeff.science at gmail.com
Tue Dec 10 15:58:55 CST 2013
Jim wanted this to go on the Wiki but I decided to inline it on the
ticket. Feel free to copy from or link to the ticket on the RMA WG
Wiki.
Jeff
---------- Forwarded message ----------
From: MPI Forum <mpi-forum at lists.mpi-forum.org>
Date: Tue, Dec 10, 2013 at 3:56 PM
Subject: Re: [MPI Forum] #398: request-based remote completion for RMA
To:
#398: request-based remote completion for RMA
-------------------------------------+-------------------------------------
Reporter: | Owner:
jhammond | jhammond
Type: New | Status: new
routine(s) | Milestone:
Priority: Not | 2013/03/11 Chicago, USA
ready / author rework | Resolution:
Version: MPI | Implementation status:
<next> | Waiting
Keywords: RMA | Author: Rich Graham: 0
Author: Bill Gropp: 0 | Author: Torsten Hoefler: 0
Author: Adam Moody: 0 | Author: Jesper Larsson Traeff: 0
Author: Dick Treumann: 0 | Author: David Solt: 0
Author: George Bosilca: 0 | Author: Rajeev Thakur: 0
Author: Bronis de Supinski: 0 | Author: Alexander Supalov: 0
Author: Jeff Squyres: 0 |
Author: Rolf Rabenseifner: 0 |
-------------------------------------+-------------------------------------
Comment (by jhammond):
Summary of working group (WG) discussion:
* The justification of fine-grain remote completion was determined to be
weak by multiple members of the WG. The two uses were stated to be CAF
2.0 and Charm++. Members of the WG could not come up with a strong
performance motivation i.e. could not come up with a hardware scenario
where separation of remote completion would measurably improve
performance.
* Brian did not like multiple requests and proposed instead to have a
single request for a new function that would have end-to-end completion
semantics, i.e. the completion of the request would correspond to
'''both''' local and remote completion - no separation of local and remote
completion is available.
* Pavan and others believe that nonblocking flush is sufficient to address
the major issue at hand, which is that blocking flush causes problems.
* Jim and others believe that returning a request from {{{IWIN_FLUSH} and
then testing on it is the wrong way to do this and proposes instead to add
{{{WIN_TEST}}}, which has different semantics than {{{IWIN_FLUSH}}}. In
particular, there is no request to be managed.
* Jeff, Brian and possibly Jim thought that there might be some value in
adding a SHMEM-like {{{PUT}}} to the MPI standard, which is a Put that
blocks on local completion so as to eliminate the need for two function
calls to achieve this effect. This feature was determined to be out-of-
scope w.r.t. this ticket and in need of its own ticket if the WG wishes to
pursue it.
--
Ticket URL: <https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/398#comment:4>
MPI Forum <https://svn.mpi-forum.org/>
MPI Forum
--
Jeff Hammond
jeff.science at gmail.com
More information about the mpiwg-rma
mailing list