[mpiwg-rma] Draft MPI RMA chapter update

[Thakur, Rajeev]

> If we are going to fix "It is often useful in a put operation to
> combine the data moved to the target process with the data that
> resides at that process, rather than replacing it."

The only change made was
"rather then (typo) replacing the data" => "rather than replacing it"

> page 46 line 10-11: "each access epoch much target a different
> process." s/much/must/

Good catch. That is a new error introduced in the editing. 

> The text added on page 47 "Completes an RMA access epoch started by a
> call to MPI_WIN_LOCK on window win." and "Completes a shared RMA
> access epoch started by a call to MPI_WIN_LOCK_ALL on window win."
> appear to add a semantic restriction not present in MPI 3.0,

The only change made was that the parameters to Win_lock(...) and lock_all(...) were omitted. They were present earlier.

Rajeev

On May 3, 2015, at 4:53 PM, Jeff Hammond <jeff.science at gmail.com> wrote:

> (1)
> 
> If we are going to fix "It is often useful in a put operation to
> combine the data moved to the target process with the data that
> resides at that process, rather than replacing it." in 11.3.4, we
> should go all the way to
> 
> "It is often useful in a remote update operation to combine the data
> moved to the target process with the data that resides at that
> process, rather than replacing it."
> 
> because "put" is not a superset of MPI_Put and MPI_Accumulate.
> 
> (2)
> 
> page 46 line 10-11: "each access epoch much target a different
> process." s/much/must/
> 
> (3)
> 
> The text added on page 47 "Completes an RMA access epoch started by a
> call to MPI_WIN_LOCK on window win." and "Completes a shared RMA
> access epoch started by a call to MPI_WIN_LOCK_ALL on window win."
> appear to add a semantic restriction not present in MPI 3.0, unless it
> says somewhere else that LOCK_ALL must be matched by UNLOCK_ALL and
> not by the appropriate number of UNLOCK calls, and vice versa.  If it
> says this, I can't remember where it is.
> 
> And if one wants to add this text, it would seem symmetric to specify
> that FLUSH_ALL and FLUSH_LOCAL_ALL should be associated with LOCK_ALL
> epochs and not N*LOCK epochs.  However, I do not think this is
> something we should do, and thus I infer we want to be less specific
> in the previous instance, meaning, we should use
> 
> "Completes an RMA access epoch started by a call to MPI_WIN_LOCK or
> MPI_WIN_LOCK_ALL on window win."
> 
> for both cases on page 47.
> 
> 
> 
> Jeff
> 
> On Sun, May 3, 2015 at 8:33 AM, William Gropp <wgropp at illinois.edu> wrote:
>> Attached is an update to the RMA chapter.  This does not introduce or change any features; it simply corrects some language and appearance (spacing) and removes a few errors in the text.  New text is in red; deletions are not shown.  Also attached is the svn diff for these items.  The chapter authors reached consensus on these; we have a list of things that the RMA group should discuss at the next meeting.  Also note, because of the short time left, the other chapter authors have not had a chance to review these - any errors are mine.
>> 
>> If you find a problem with these updates (not just awkward wording, or a new problem), respond to the whole group so that we can discuss it as quickly as possible.
>> 
>> Bill
>> 
>> 
>> _______________________________________________
>> mpiwg-rma mailing list
>> mpiwg-rma at lists.mpi-forum.org
>> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-rma
> 
> 
> 
> -- 
> Jeff Hammond
> jeff.science at gmail.com
> http://jeffhammond.github.io/
> _______________________________________________
> mpiwg-rma mailing list
> mpiwg-rma at lists.mpi-forum.org
> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-rma