I think this is just our fundamental disagreement: I believe that splitting up parts of a single request will work against you in terms of performance and scalability. The only good reason I can see to segment things is in order to isolate bad parts that you intend to leave under-powered and slow. Splitting up a single request doesn't work, since all the parts have to finish before the response can be sent, i.e. the response can only be as fast as the slowest part.

As far as SOAP/XML-RPC is concerned, they can be useful for interop with systems in other languages but they are not an efficient way to distribute load within a single system. They have serious overhead, with all the extra XML creating/parsing and network traffic.