(j3.2006) [WG9] Detailed Agenda for the next meeting of WG 9 on Friday, October 16, 2015 starting at 9:00 AM EDT in Bennington, VT

[Van Snyder]

Miguel:

In my original message, I should have referred your attention also to an
extension of the coarray features of Fortran 2008, implemented as a
Technical Specification to be incorporated into Fortran 2015.  Look for
document N2074 at http://www.nag.co.uk/sc22wg5/docs.html

Best regards,
Van

On Thu, 2015-10-15 at 07:35 -0700, Van Snyder wrote:
> On Thu, 2015-10-15 at 13:28 +0000, Luis Miguel Pinho wrote:
> > Dear Van Snyder
> > 
> > Thank you very much for the pointers. 
> > 
> > If I am understanding correctly the notion of coarray, then each
> > program image gets part of the matrix, in its coarray. 
> > 
> > I think this is similar to the notion of a view of Cilk's
> > hyperobjects, but for matrixes, which is something we are currently
> > looking at.  
> > 
> > BTW, I could not understand if there is any synchronization. Can an
> > image access the coarrays in other images? If so, what would be the
> > value?
> 
> Miguel:
> 
> An image can access coarrays in other images, as if the array were
> spread across all images, by appending what are called cosubscripts in
> square brackets.  Suppose a variable named A is a coarray that is an
> array on each image.  For simplicity, assume it has one dimension on
> each image, and that the images are arranged in a one-dimensional form.
> In Fortran terminology, this would be called a variable of rank one and
> corank one.  To access the i-th element on the j-th image, one would
> write "A(i)[j]".  This can appear in either a variable-reference or
> variable-definition context.  If a coarray is referenced without
> cosubscripts, the reference is implicitly within the executing image.
> 
> Coarrays can be scalars on each image.  The current limit in the Fortran
> standard is that rank+corank is less than 16.
> 
> One can write in one statement a computation that requires numerous MPI
> references.  Consider, for example, a halo exchange in a solver for an
> elliptic PDE.
> 
> There are synchronization statements that divide the execution on each
> image into segments, and impose relative ordering between segments on
> different images.  There are rules about referencing coarray values on
> one image that have changed on a different image.  The short story is
> that this is prohibited in unordered segments, but there are also
> primitives for finer-grained synchronization, involving locks and
> events.
> 
> The number of images is specified by the environment that launches the
> program, not by statements within the program.  Intrinsic functions are
> provided to inquire the number of images, and for each image to inquire
> its image number.  A technical specification shortly to be published,
> and that will be incorporated into the Fortran 2015 standard, will add
> the ability to divide the images into teams.
> 
> Toon Moene, one of the delegates to the ISO Fortran committee, reported
> that a substantial part of one weather model used by the ECMWF was
> converted from MPI to coarrays, and thereby ran substantially faster.
> Since the code became shorter and clearer, their ongoing maintenance
> cost is also decreased.
> 
> I'm happy to answer more questions, and to put you into contact with the
> delegate from Cray who did most of the standards-development work.  Cray
> has had coarrays for about twenty years, but they are just now making
> their way into the Fortran standard.  Cray has a C++ implementation.
> There is an effort afoot to develop an extension of C called "Uniform
> Parallel C" or UPC, based upon coarray ideas, but the work is very
> primitive (i.e., not the quality necessary for an ISO standard), and it
> will probably never be incorporated into the C standard.
> 
> Best regards,
> Van
> 
> > 
> > Best regards
> > 
> > Miguel
> > 
> > > -----Original Message-----
> > > From: Van Snyder [mailto:van.snyder at jpl.nasa.gov]
> > > Sent: 13 de outubro de 2015 21:02
> > > To: Luis Miguel Pinho <lmp at isep.ipp.pt>
> > > Subject: Re: [WG9] Detailed Agenda for the next meeting of WG 9 on Friday,
> > > October 16, 2015 starting at 9:00 AM EDT in Bennington, VT
> > > 
> > > Luis:
> > > 
> > > Concerning parallelism, you (and perhaps other WG9 members) might be
> > > interested in the coarray features added to Fortran in 2008.  This is an SPMD
> > > system that uses an economical syntax.  Search for "coarray" in the current
> > > draft for the next revision of Fortran, informally called "Fortran 2015." at
> > > http://j3-fortran.org/doc/meeting/208/15-007r2.pdf.
> > > 
> > > There is a summary of coarrays, as of Fortran 2008, at
> > > http://www.nag.co.uk/sc22wg5/docs.html.  Look for document N1824.  This
> > > does not include features to be added in the next revision.
> > > 
> > > Best regards,
> > > Van Snyder
> > > 
> > > On Tue, 2015-10-13 at 18:14 +0000, Luis Miguel Pinho wrote:
> > > > Hi Joyce
> > > >
> > > >
> > > >
> > > > Unfortunately I will not be able to attend, even in webex, as I have
> > > > teaching duties during all afternoon.
> > > >
> > > >
> > > >
> > > > Report is as usual:
> > > >
> > > >
> > > >
> > > > There is no information on the restructuring of the standardization
> > > > activities in Portugal. Portugal is a P member of SC22, but AFAIK WG9
> > > > is the only WG active in the SC. National work related to WG9
> > > > continues to be in (rare) direct connection with the NB.
> > > >
> > > > The main focus of the technical work being developed continues to be
> > > > in the parallelism extensions being proposed for Ada.
> > > >
> > > >
> > > >
> > > >
> > > >
> > > > Best regards
> > > >
> > > >
> > > >
> > > > Miguel
> > > >
> > > >
> > > >
> > > >
> > > >
> > > >
> > > >
> > > >
> > > >
> > > > From: Dr. Joyce L Tokar [mailto:tokar at pyrrhusoft.com]
> > > > Sent: 20 de setembro de 2015 20:22
> > > > To: 'WG9 List' <wg9 at ada-auth.org>
> > > > Subject: [WG9] Detailed Agenda for the next meeting of WG 9 on Friday,
> > > > October 16, 2015 starting at 9:00 AM EDT in Bennington, VT
> > > >
> > > >
> > > >
> > > >
> > > > Attached please find the detailed agenda for the next meeting of WG 9.
> > > > For those of you planning to attend the meeting virtually, the dial-in
> > > > information is in this announcement.
> > > >
> > > >
> > > >
> > > > Project editors, liaisons, and any National Body expert:  Please
> > > > forward me you reports, no later than Tuesday, October 13, 2015.
> > > >
> > > > Similarly, please notify me of any persons wishing to attend this
> > > > meeting as a guest, no later than Tuesday, October 13, 2015.
> > > >
> > > >
> > > >
> > > > Thank You,
> > > >
> > > > Joyce
> > > >
> > > >
> > > >
> > > > Joyce L Tokar, PhD
> > > >
> > > > Pyrrhus Software, LLC
> > > >
> > > >
> > > >
> > > > ISO/IEC JTC 1/SC 22/WG 9 Convenor
> > > >
> > > >
> > > >
> > > >
> > > >
> > > >
> > > 
> > 
>