[openib-general] Re: RFC userspace / MPI multicast support

[Sean Hefty]

amith rajith mamidala wrote:
> 1. Does the API which waits for join to complete
> ensure that the multicast forwarding tables in the switches have been
> updated. This is one of the main problems that we had studied:

The join is asynchronous.  Completion of the join would not be reported until 
the switch tables had been updated.  (Note that this is really an SA issue 
outside the scope of the local code).

> 2. I am not clear on how to access the QP associated with the cm_id for
> multicast. This includes posting the receive descriptors etc.

rdma_create_qp() returns a struct ibv_qp *.  You would access the QP just as you 
normally would for posting receives or sends to non-multicast endpoints.

> 3. If an multicast address is already used by an application running on
> the cluster and if another request is made by a different application with
> the same multicast address, does this generate an error? From the API, it
> looks like the application has to manage this aspect,

If two applications make a request to join the same group, they simply both join 
the same group.  One of the key points to the proposal is that a multicast group 
is identified by an IP address from a user's perspective.  MCMemberRecords are 
abstracted from the user.

- Sean