[Mpi-forum] MPI-3 pre-proposal

[Jeff Squyres]

On Apr 14, 2008, at 5:18 PM, Solt, David George wrote:
> We just use lots and lots of environment variables for this  
> communication.  :)

Ditto.  :-)

> We have had several requests for a mechanism to allow the  
> application to get some feedback from MPI of which interconnect got  
> selected.

This might get a bit tricky to standardize because the answer differs  
between each pair of peers -- you could need up to 0.5*(N^2) pieces of  
information (although some heuristics could be applied to reduce all  
links that are the same).  How to represent this in a string that is  
assumedly parsable by the MPI app would be... er... interesting.  It's  
worth talking about, IMHO.  :-)

> Some mechanism to query this and some defecto standardization on  
> what MPI's reported would probably be helpful to many ISVs.

Agreed.

> Of course we can use some MPI_Info caching on MPI_COMM_WORLD if it  
> is sufficient to get this information after MPI_Init().   I think we  
> had at least one request to provide this information before  
> MPI_Init().

FWIW: I was hoping to avoid new standardized attribute keys because a)  
it may limit what kind of information can be used/provided by an MPI  
implementation, and b) attribute keys are only valid on 3 types of MPI  
handles.  I don't think you were trying to say that, but I thought I'd  
put out this bias because I didn't include it in my first mail.

-- 
Jeff Squyres
Cisco Systems