[Mpi3-rma] Info accumulate_ops
rabenseifner at hlrs.de
Sat Oct 20 06:27:49 CDT 2012
> ... So a valid user program in MPI-2.2 was only
> allowed to use the same OP for concurrent accumulates. ...
I was never aware of this restriction.
But you are right: MPI-2.2 page 365 lines 31-33:
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
predened datatype, on the same window.
----- Original Message -----
> From: "Torsten Hoefler" <htor at illinois.edu>
> To: "MPI 3.0 Remote Memory Access working group" <mpi3-rma at lists.mpi-forum.org>
> Sent: Friday, October 19, 2012 8:38:21 PM
> Subject: Re: [Mpi3-rma] Info accumulate_ops
> On Tue, Oct 16, 2012 at 01:35:08PM +0000, Hubert Ritzdorf wrote:
> > 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"
> This would be illegal (as described in the semantics section). The
> was an exception to allow querying a flag with our "fake atomic reads"
> while performing atomic updates CAS etc.. So really only the two
> mentioned combinations are legal.
> > (*) 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".
> Default value means that this is the default semantic if the key is
> set (or not evaluated). So a valid user program in MPI-2.2 was only
> allowed to use the same OP for concurrent accumulates. In MPI-3.0 it
> use the same op and also NO_OP in addition. Thus, this richer semantic
> cannot break backward compatibility. Setting the info key enables the
> user to enable the more limited MPI-2.2 semantics which may unlock a
> of richer optimizations possible in the library.
> ### qreharg rug ebs fv crryF ------------- http://www.unixer.de/ -----
> Torsten Hoefler | Assistant Professor
> Dept. of Computer Science | ETH Zürich
> Universitätsstrasse 6 | Zurich-8092, Switzerland
> CAB E 64.1 | Phone: +41 76 309 79 29
> mpi3-rma mailing list
> mpi3-rma at lists.mpi-forum.org
Dr. Rolf Rabenseifner . . . . . . . . . .. email rabenseifner at hlrs.de
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)
More information about the mpiwg-rma