[mpich-discuss] MPICH2 & Large dat

[Rob Ross]

If you are particularly worried about portability, you could build a  
contig of some number of MPI_CHARs (e.g., 4 to continue the int  
example), then provide that type as the input to the routine, with a  
correspondingly smaller count.

Rob

On Oct 2, 2008, at 9:00 AM, Nicolas Rosner wrote:

> Yes, but what he is suggesting is that you could "lie" to MPI about
> the data type.  For example, if you wanted to send 1000 MPI_CHARs, you
> could instead send them as if they were 250 MPI_INTs (assuming an int
> takes up as much space as four chars on your system).  Since 250 is a
> smaller count than 1000, this could be of use to you (because your
> problem is that the count is becoming too big).
>
> Of course, you need to be extra careful if you try this, for it could
> affect portability (after all, the whole point of having symbolic
> datatypes like MPI_INT for in-flight data is to avoid having to
> remember, or even assume, that an int takes up four bytes).
>
> HTH,
> Nicolás
>
>
> On 10/2/08, Hisham Adel <hosham2004 at yahoo.com> wrote:
>> The problem is in the count number not in the data type.
>>
>> Thanks  a lot
>>
>>
>>
>> --- On Thu, 10/2/08, The Source <thesourcehim at gmail.com> wrote:
>> From: The Source <thesourcehim at gmail.com>
>> Subject: Re: [mpich-discuss] MPICH2 & Large data
>> To: mpich-discuss at mcs.anl.gov
>> Date: Thursday, October 2, 2008, 3:08 PM
>>
>> Send, using different MPI data types. For example MPI_INT instead of
>> MPI_CHAR, so element number will be 4 time smaller. The other way  
>> is to
>> split data to several messages.
>