[Mpi3-rma] Urgent: Target displacement sign issue

[Jim Dinan]

Hi Sayantan,

This will require more discussion, so we'll have to handle it in the MPI 
3.0 errata.

The proposed solution should not impact communication, it is only to 
enable full addresses (displacements in dynamic windows) to be passed as 
the displacement argument to get/put/acc/etc without generating an error.

  ~Jim.

On 9/12/12 4:17 PM, Sur, Sayantan wrote:
> Hi Jim,
>
> Wondering what happened to this issue. I see that Rich sent a message saying the final version has been uploaded.
>
> Anyways, I'm not sure if this would have qualified as a "ticket-0" change. While the solution proposed is portable, this requires the target to communicate with the origin. For PGAS languages, this would mean probably broadcasting this address to every other process in the job. Alternatively, this could be done in an on-demand basis upon first PGAS put or get.
>
> Not that it is a big deal, just a significant change to come late. MPI-next?
>
> Thanks.
>
> ===
> Sayantan Sur, Ph.D.
> Intel Corp.
>
>> -----Original Message-----
>> From: mpi3-rma-bounces at lists.mpi-forum.org [mailto:mpi3-rma-
>> bounces at lists.mpi-forum.org] On Behalf Of Jim Dinan
>> Sent: Wednesday, September 12, 2012 8:36 AM
>> To: mpi3-rma at lists.mpi-forum.org
>> Subject: [Mpi3-rma] Urgent: Target displacement sign issue
>>
>> Hi All,
>>
>> We're down to the wire on this change.  As specified, dynamic windows will
>> /not/ work reliably in MPI 3.0.  We currently see "negative"
>> displacements for dynamic windows in MPICH on 32-bit platforms.
>> Earlier, I suggested a change stating that the implementation will ignore the
>> sign for dynamic windows.  Here's a revised proposal.
>>
>> In the paragraph on page 425, line 1, add the bracketed sentence:
>>
>> In the case of a window created with MPI_WIN_CREATE_DYNAMIC, the
>> target_disp for all RMA functions is the address at the target; i.e., the
>> effective window_base is MPI_BOTTOM and the disp_unit is one. [Thus, for
>> dynamic windows, the target_disp argument to RMA communication
>> operations is not restricted to non-negative values.] Users should use
>> MPI_GET_ADDRESS at the target process to determine the address of a
>> target memory location and communicate this address to the origin process.
>>
>>    ~Jim.
>> _______________________________________________
>> mpi3-rma mailing list
>> mpi3-rma at lists.mpi-forum.org
>> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma
>
> _______________________________________________
> mpi3-rma mailing list
> mpi3-rma at lists.mpi-forum.org
> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma
>