The best practice is to have a VM profile that goes to the Connection subscriber first, then the publisher.  It would be possible to load balance some with two VM profiles and this would work, but it's probably not worth the trouble.  Since Connection only supports a publisher and single subscriber, and the system must be sized such that each individual server can handle the full call load. 

Incidentally, in normal circumstances, the primary Connection server is the publisher.  This means that all notifications will always be sent from this server.  So even though the calls to Connection would arrive on the subscriber, the publisher is still busy doing notifications as well as delivering the voicemail messages themselves.

On the Connection server, the mailbox/subscriber data is present on both systems.  They are administered as one, so there's no way to have a subscriber on one server and not the other.