[Mpi3-tools] MPI handles interface

Jeff Squyres jsquyres at cisco.com
Fri Apr 17 07:11:42 CDT 2009

On Apr 16, 2009, at 7:30 PM, Bronis R. de Supinski wrote:

> Hmm. My expectation was there would be query functions
> for individual fields, a la MPI_Get_count. I guess I
> don't have a strong feeling about it -- when you want
> only one field, you will pay a somewhat significant
> cost with your approach but your approach will be an
> even bigger win when you want all of them. It may be
> harder for an implementation to add fields with your
> approach, I don't know...

I could certainly make an accessor function for *each* field.  But  
instead I grouped them into "related" fields.  E.g., if you're going  
to ask for the how many dimensions there are in the cartesian  
topology, you're probably also going to want to the array of the  
dimension lengths as well (at least, that's what I assumed).

This got me down to 6 query functions instead of 21, which, all things  
being equal, I figured was a good thing.  I certainly could make them  
a bit more fined-grained -- for example, the initial function  
(_query_basic()) is returning a bunch of basic information that isn't  
really grouped together:

- communicator name
- bit flags
- this proc's rank
- size
- fortran handle
- c++ handle

If people feel strongly, I'm happy to make 21 query functions.   
Whateveah.  :-)

As for adding implementation-specific fields, I don't think this  
proposal addresses that at all.  If Microsoft wants to make a proposal  
about adding vendor-specific fields, I'd love to see it.  :-)

Jeff Squyres
Cisco Systems

More information about the mpiwg-tools mailing list