Implement MPI_Comm_set_info() and MPI_Comm_get_info()

[gpaulsen]

Right now MPI_Comm_set_info() and get_info() are dummy functions that just throw away whatever data is given to it, and doesn't return anything.
These need to be implemented to work as stated in the MPI 3.0 Standard.

Dr. David Solt (IBM) has agreed to implement this functionality in the March 2015 timeframe.

[jsquyres]

We also talked about separating the functionality into two parts:

1. Move the central "key=value" functionality stuff down to the OPAL layer.
2. Keep the MPI semantics (implemented on top of the OPAL "key=value" stuff) up in the OMPI layer.

This would allow us to take values from the app that were set via MPI_Info API calls and pass them down to OPAL entities -- e.g., BTLs.

What's Dave Solt's Github ID?

[gpaulsen]

I pinged Dave, and he thinks he might be able to start tomorrow.

[dsolt]

Tomorrow, tomorrow I'll love you tomorrow. Its only a day away. Ok, I'm seriously downloading the code and working on it right now!

[dsolt]

Can you think of a good MPI_Info that we could add a "subscriber" for? I could make one up? Like buffering=none to turn all MPI_Send's into MPI_Ssends?

[jsquyres]

Sounds good to me. :-)

[dsolt]

Does MPI_XXXX_get_info return an Info with all key/values that might be used by the underlying system even if the values are garbage, or only return key/values for which the underlying system has validated that the value made sense and had some influence on the system. I plan to go with the later unless you think otherwise.

[edgargabriel]

Not sure whether it makes sense in this context or not, a looong time back (pre Open MPI) we had that idea of trying to separate communication on a per communicator basis, e.g. all communication on comm1 uses InfiniBand (or ATM at that time), on comm2 GE, or similar. Today, we could think about trying to separate communication for File I/O and 'message passing' to use separate IB cards or ports (if you have multiple per node). An Info object might be optimal for that.

[jsquyres]

MPI 3.1 section 6.4 p249-250 says:

MPI_COMM_GET_INFO returns a new info object containing the hints of the commu- nicator associated with comm. The current setting of all hints actually used by the system related to this communicator is returned in info_used. If no such hints exist, a handle to a newly created info object is returned that contains no key/value pair. The user is responsible for freeing info_used via MPI_INFO_FREE.
Advice to users. The info object returned in info_used will contain all hints currently active for this communicator. This set of hints may be greater or smaller than the set of hints specified when the communicator was created, as the system may not recognize some hints set by the user, and may recognize other hints that the user has not set. (End of advice to users.)