ISO/IEC JTC1/SC22/WG5 N1906

        Response to the second invitation to comment on the 
          contents of the TS on further coarray features

                           John Reid

                        7 February 2012


In N1888, I invited comments on the proposals there or new suggestions.
I have received only one response. This was from Reinhold Bader. Here 
it is.


Comments on WG5/N1888:
~~~~~~~~~~~~~~~~~~~~~~

Comment on Proposal 2 (teams):

Perhaps it should also be mentioned that teams, if properly designed,
would provide Fortran with a uniformly usable hybrid programming model, 
well-mapped to nowadays' deeply hierarchical HPC architectures;
synchronization statements would not only not impede uninvolved images, 
but might also execute much faster within a team (e.g., executing an 
OpenMP lock across a large-scale shared memory system typically needs 
the top-level latency of the system, multiplied with the logarithm 
of the number of units on that level). 
Adding a WITH TEAM construct therefore may turn out to be more than 
only syntactic sugar. 

Comment on Proposal 5 (global pointers):

The discussion section mentions that "[t]he Bader proposal is
significantly more complicated since it also contains the idea of
any image being able to access a local object that resides on one
image". This appears to be a misunderstanding - coscalar pointers
were certainly not meant to be able to point at local entities
on other images. Although I may not have explicitly said this, I 
did mention that the intent was to mirror the functionality of 
UPC shared pointers to shared entities. (Note that I am not 
arguing against the conclusion in the proposal, although I suspect 
it increases the size of the feature). 


Comment on Proposal 11:

For a decision on the size of the TS, priorization may be helpful. 
My preference for priorization would be as follows, in descending 
order of importance:

(1) Collective procedures
(2) Notify/Query one-sided synchronization
(3) Teams
(4) Additional atomic operations
(5) Remove restrictions on coarray components
(6) Global pointers
(7) Parallel I/O extensions
(8) Predicated copy_async intrinsic

The order of importance was chosen mainly according to my expectation
on the combined impact on performance scalability and ease of use for
the programmer.

Even if a 30% size increase in the time budget is conceded, I suspect 
that not all the suggested features still open for investigation can 
be realized.