[openib-general] RFC userspace / MPI multicast support

[Sean Hefty]

Matt Leininger wrote:
>   Requiring IPoIB to be running for MPI to work is a new dependency that
> users are not accustomed to.  Maybe we need to do this through the cm
> rather than rdma_cm.  We shouldn't require IPoIB to be running for MPI
> to use IB.

Use of these new features already carries with it new dependencies, in the form 
of new libraries and kernel modules.

Are MPI users even accustomed to having the IB CM loaded?

How does MPI discover the remote nodes in the fabric today?  How does it obtain 
the path records?  How does it use multicast groups today?

I'm trying to find a solution that can be both easy to use, yet allow for low 
level control.  However, I don't necessarily want to restrict an implementation 
based on existing applications that are either hard-coded or coded for static 
configurations.

- Sean