(j3.2006) synchronizing I/O

[Bill Long]

We discussed this topic in plenary this morning.   A new approach 
emerged that transfers the work from the standard to the vendor, more or 
less.  The approach is to keep stdout and stderr as separate "files" in 
each image from Fortran's point of view.  And then include a Note that 
the processor might be expected to interleave the records of all these 
files onto the user's terminal in some reasonable way.  Van pointed out 
that this already occurs for the combination of stdout and stderr, both 
of which are (normally) directed to the terminal with interleaved 
records.  This removes the exceptional nature of these files from the 
standard (a simplification), but does not mandate the user's 
expectations.  The DATA subgroup will write a paper to implement this 
approach.

Cheers,
Bill



Aleksandar Donev wrote:
> On Tuesday 10 February 2009 04:01, John Reid wrote:
>   
>> You begin to convince me that we should require the user to ensure
>> that there is an ordering of the i/o statements to one of these
>> files.
>>     
> What would the requirement mean:
> 1) The user can write to stdout from unordered segments, but there is no 
> guarantees that what comes out is not garbled (this is pretty much MPIs 
> no-guarantee approach).
> 2) Two WRITEs to stdout by two different images must be in ordered 
> segments
> I can go along with the first. If you put in the second, then it makes 
> it harder for vendors to do extensions and allow nice and quick sorted 
> I/O. After all, it isn't that hard to implement---just buffer until the 
> record is complete, no?
> Aleks
>
>   

-- 
Bill Long                                   longb at cray.com
Fortran Technical Support    &              voice: 651-605-9024
Bioinformatics Software Development         fax:   651-605-9142
Cray Inc., 1340 Mendota Heights Rd., Mendota Heights, MN, 55120