07-164r1

To:         J3
From:       Bill Long
Subject:    Co-arrays and processor dependent order
Date: 2007 February 15
References: J3/07-007, 06-342, 06-278, 06-279

This is a re-submittal of 06-342 which was deferred at meeting
178.

These comments are gleaned from 06-278 and 06-279 which were
deferred at meeting 177.

Commment:

Page 73:20 says the order in which things are finalized is
processor dependent. Page[131:22-24] also mention automatic
operations without an order. Does processor dependent allow
different images to do things differently?  Can
different images finalize things in different orders? I don’t
know if it matters in the co-array sense; but I don’t think we
say which it should be anyplace.

Reply:

Within an image, the order of finalization is processor-dependent, but
each image should execute the finalizations in the same order.

Edits:

[73:15-16] In 4.5.6.2 The finalization process, list item (2): replace
"processor-dependent" with "processor-dependent and
image-independent".

[73:21] Replace "processor-dependent" with "processor dependent and
image independent".

-------------------

Comment:

Are co-arrays expected to work on different kinds of hardware
in a network? It’s maybe reasonable to expect wildly different
hardware and OSs to have different memory management methods
and, maybe, to link things in different orders and do the
deallocate, or whatever, in a different order. Is this allowed
or prohibited?

Either way, I think some words in 2.3.2 would help.

For what it’s worth, many of the clause 13 collective
Functions say something like "processor-dependent and image-
dependent..."

Reply:

It is not prohibited. However, an implementation for such an
environment would be significantly more complicated than for a
homogeneous system.

Edit:

[14:24+] In 2.3.3 Program execution, at the end of Note 2.4, add
"Though it is simpler to implement co-arrays on a set of identical
CPUs, such a configuration is not required."


---------------

Comment:

For [130:15] Malcolm suggests saying it in only the correct
place and replacing the text with
"Deallocation of allocatable components can also occur during
intrinsic assignment (7.4.1.3)."

Reply:

The proposed replacement text appears to be unrelated to [130:15], so
we have no reply for this question.

---------------