(j3.2006) (SC22WG5.3644) [ukfortran] A comment on John Wallin's comments on Nick MacLaren's comments

[N.M. Maclaren]

On Nov 6 2008, Bill Long wrote:
>
>> Most of the users I have dealt with have backed off shared-memory
>> paradigms when they found that they couldn't debug or tune them, and
>> gone back to MPI.  The problem is that there are, and can be, no tools
>> to trap race conditions.
>
>This is, of course, one of the main arguments in favor of coarrays.  The 
>programming model, is SPMD,  as with MPI, which experience has shown to 
>work, and to be the most popular.   Shared-memory models, like OpenMP, 
>do have their place, but have limitations. 

Not at all.  Coarrays are almost indistinguishable from OpenMP in this
context.

The debugging problems were caused by race conditions in shared memory
accesses, which are impossible in MPI if you avoid using non-blocking
transfers in open code.  And the killer was the lack of any tools to
detect them.  Both apply to coarrays.

The tuning problems were caused by similar conditions, but with false
sharing replacing race conditions, and the inability to put instrumentation
into the communication.  Again, the same is true for coarrays.

>Deadlocks introduced by the implementation should be reported as bugs to 
>the vendor. 

Unless the standard makes it clear whether the programmer or implementor
is at fault, that is merely a waste of time.  MPI goes to great trouble
to specify what each of them is required to do, and that is a major reason
for its reliability and portability.


Regards,
Nick Maclaren,
University of Cambridge Computing Service,
New Museums Site, Pembroke Street, Cambridge CB2 3QH, England.
Email:  nmm1 at cam.ac.uk
Tel.:  +44 1223 334761    Fax:  +44 1223 334679