We have an environment that is expected to scale out to 200,000 queues representing mobile devices.  Preliminary testing indicates that performance begins to fall off at around 50,000 to 75,000 addresses with a modest load of around 2,000,000 messages per hour on a 12-core system with around 32GB of heap.  Very preliminary investigation suggests that the bottleneck may be with the netty thread finding matching bindings for the messages.  Horizontal scaling would be difficult in this environment and we would like to push the limit of vertical scaling if possible.