The motivation for a ``content layer'' approach came as part of the TRIAD[2] project. TRIAD is a new, NAT-friendly Internet architecture which seeks to reduce dependency on addresses by promoting names as transport-layer endpoints. In a TRIAD Internet, all large-scale routing would occur at the naming level. We believe this approach is ultimately more scalable and deployable than attempts to solve problems (such as mobility, multihoming, anycasting, and wide-area addressing) at the network level.
Two features of TRIAD enhance the content routing architecture. TRIAD provides extended addressing via the Wide-Area Relay Addressing Protocol (WRAP), which provides loose-source routing among multiple address realms. WRAP addresses can be used to specify a path through the network, ensuring that the route selected by the content routing layer is the path actually used by data packets. TRIAD also integrates TCP connection setup into INRP name lookup; by sending TCP connection initiation information inside an INRP request, the latency for web transactions can be lowered yet further. Thus, the full TRIAD architecture integrates naming, routing, and connection setup into a single framework.
INRP allows proxies and web caches to intercept content requests based on URL. We have not implemented or fully explored this design, but it appears to be a promising way to provide ``semi-transparent'' proxies, which would require no explicit configuration at the client, but would be used by the client as a content request's TCP connection endpoint.
Finally, the integration of naming and routing allows feedback-based routing. Conventional IP routing schemes have few ways to tell if the routes they select actually deliver packets to the intended destination. Content routers, however, can track the responses they receive to forwarded queries, allowing them to make better decisions and react more quickly to routing problems than conventional routers. For example, if content servers send back load information in INRP responses, then content routers can obtain up-to-date load information on heavily used sites without placing this load information into routing updates.
The most current version of this paper can be found at [9].