[MPI3 Fortran] Proposing changes to Fortran 2008

Jeff Squyres jsquyres at cisco.com
Thu Mar 20 14:26:59 CDT 2008


On homogeneous architectures -- where the same hardware and software  
is used on all sides (to include the compiler) -- yes, this will  
likely work just fine.  And probably most people do this, because it  
can be quite a bit easier.

If your data is dense in memory, that's probably "good enough". If  
your data is sparse, you could end up sending quite a bit of garbage  
along with your real data, decreasing the bandwidth available to your  
app, potentially adding more latency, etc.

In heterogeneous situations (where any of the above assumptions is not  
true), it clearly may not work properly.




On Mar 20, 2008, at 1:47 PM, Aleksandar Donev wrote:
> On Thursday 20 March 2008 10:35, Jeff Squyres wrote:
>> because the
>> user application will have had to created an MPI datatype *at run
>> time* that will have accurate offsets/types of each member.
> Out of curiosity: Have you seen codes that actually do that?
> What I would do, to make my life simple, is treated the whole scalar  
> of
> derived type as a blob of memory and ignored what fields are  
> actually inside,
> in what order, and with what padding. On homogeneous architectures  
> doesn't
> this work just fine?
> Aleks
>
> _______________________________________________
> mpi3-fortran mailing list
> mpi3-fortran at lists.mpi-forum.org
> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-fortran


-- 
Jeff Squyres
Cisco Systems




More information about the mpiwg-fortran mailing list