OK, as I remember it (I'll have to go back and read my own book on this one!) we do have some localization of requests from the EJB client. Is what you're seeing that you create one client stub, send a bunch of messages from it to the servers, and they all go to the same server? If so, that's normal.
On the other hand, if you create multiple client stubs in either the same program or in different programs then
you should see the requests from the different stubs spread across both servers.
Also, if you bring one server down you should see transparent failover to the other server as well in any case.
Kyle