The proxy is probably a good idea but I don't think the problem is as hard as it sounds, especially if you can relax the latency and/or message ordering requirements. If you build a hierarchy of queues and are smart about using reference counts you won't have nearly as many reads and writes as the article indicates.

Tibco and others have been doing this sort of thing for 20 years (only they call it Subject Based Addressing) for pretty demanding customers.