I had raised this issue during the chapter review. It felt odd to me that the default was for some optimized behavior which can potentially lead to erroneous results in case the program doesn't obey this behavior. MPI-3: "This can eliminate the need to protect access for certain operation types where the hardware can guarantee atomicity." It appears MPI-2 had the same restriction on concurrent ops, just not through an info key though. MPI-2.2, 11.7, Pg 365, line 29-34: "A location in a window must not be accessed as a target of an RMA operation once an update to that location has started, until the update becomes visible in the public window copy. There is one exception to this rule, in the case where the same variable is updated by two concurrent accumulates that use the same operation, with the same predefined datatype, on the same window." === Sayantan Sur, Ph.D. Intel Corp.
-----Original Message----- From: [email protected] [mailto:mpi3-rma- [email protected]] On Behalf Of Rolf Rabenseifner Sent: Friday, October 19, 2012 7:25 AM To: MPI 3.0 Remote Memory Access working group Subject: Re: [Mpi3-rma] Info accumulate_ops
I would like to see an errata proposal from the RMA group. - It is unclear, what does it mean, not to use the info accumulate_ops. Info keywords cannot have a default value because you must define the keyword always together with a value. - It is unclear which value would be returned with info_get if an MPI library handles this keyword, but the application has not set it. - It is unclear which behavior we have when an MPI library does not support this info: Is it MPI-2.2 behavior?
Best regards Rolf
----- Original Message -----
From: "Hubert Ritzdorf" <[email protected]> To: [email protected] Sent: Tuesday, October 16, 2012 3:35:08 PM Subject: [Mpi3-rma] Info accumulate_ops Hi,
I think that there is a problem in the definition of info argument "accumulate_ops" of MPI_Win_create functions (Page 406, Line 4-9 of MPI 3.0 document).
(*) A value is missing if not all concurrent accumulate calls perform the same operation. Possible values might be: "not_same_op" or "false" (*) The actual default value of "same_op_no_op" breaks the backward compatibility and causes problems in portability of applications, since a MPI implementation is not required to evaluate (recognize) info keys. Therefore, "not_same_op" must be the default value of info argument "accumulate_ops".
Hubert
_______________________________________________ mpi3-rma mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma
-- Dr. Rolf Rabenseifner . . . . . . . . . .. email [email protected] High Performance Computing Center (HLRS) . phone ++49(0)711/685-65530 University of Stuttgart . . . . . . . . .. fax ++49(0)711 / 685-65832 Head of Dpmt Parallel Computing . . . www.hlrs.de/people/rabenseifner Nobelstr. 19, D- 70550 Stuttgart, Germany . (Office: Allmandring 30) _______________________________________________ mpi3-rma mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma