[mpich-discuss] romio large array

[Dave Goodell]

On Apr 26, 2010, at 9:38 AM, Rob Latham wrote:

> On Mon, Apr 26, 2010 at 08:05:41AM -0500, Dave Goodell wrote:
>> [adding RobL]
>>
>> It looks like ROMIO is checking for reads of more data than INT_MAX
>> bytes in a single call.  I'm guessing that this is defensive because
>> some of the internal code in ROMIO may not be prepared to handle
>> this case.  Rob can probably comment with less speculation.
>
> that's exactly what's going on.

This is a stricter check than what the standard implies.  ROMIO is  
checking that ((type_size*count)<=INT_MAX).  It really should only  
need to check that (count<=INT_MAX) as long as all calculations are  
internally handled with a larger (64-bit) type.  This would actually  
permit the read in this user's instance.  Would it be hard to move to  
the latter case?

> Dave, I don't know if the forum is still talking about the MPI_COUNT
> issue, but we get questions like this about once a year. People
> definitely try to read more than 2GB of data at a time.

It is under discussion.  But AFAIK all of the discussion is about how  
to fix the "int count" parameters.

-Dave