Tomasz Janczuk wrote a great series of articles on the HTTP Polling Duplex protocol. Specifically, you might want to look at this one which discusses scalability directly.
The gist of it is that in order to provide scalability, you have to support the Duplex Service with a backend bus architecture. The services would be responsible just for shuttling messages from the client to the backend. The backend would be responsible for identifying where outgoing messages should be routed and how to process inbound messages.