There has been increasing research effort in making the Internet architecture capable of providing different service levels to meet QoS requirements. Integrated Services (Int-Serv) with RSVP signaling introduces end-to-end per-flow reservations, which requires core routers to maintain individual flow states and may not scale well. Differentiated Services (Diff-Serv), on the other hand, relies on packet markers and per-hop behavior (PHB) at core routers to provide service differentiation. Diff-Serv bandwidth brokers (DSBBs) are proposed to negotiate resource reservations between neighboring domains. However, admission control is only performed at the ingress edge router to a Diff-Serv domain, and reservations are set up for aggregate traffic on hop-by-hop basis without any global knowledge of the network conditions. Diff-Serv trades off end-to-end QoS for scalability.

The lack of a well-studied policy architecture to regulate resource provisioning in a scalable manner has motivated our design of a Clearing House (CH) as an alternative solution. The CH attempts to provide better QoS assurance and higher network utilization, as offered by stateful networks (e.g., Int-Serv), while maintaining the scalability of a stateless network architecture (e.g., Diff-Serv). We also address two other problems that are often encountered today: (a) lack of a methodology for intra- and inter-domain provisioning or service composition based on dynamic traffic patterns, and (b) lack of distributed resource controlling systems for large domains. DSBBs regulate inter-domain reservations using a flat architecture with one broker per domain, regardless of the size and population of the domains. Our initial simulation results show that such architecture does not scale for large domains, and causes a single point of congestion.

Our CH architecture uses a hybrid of a flat and a hierarchical structure. A hierarchical structure helps in distributing the network state information among the various CH-nodes and reduces the amount of states maintained, while a flat structure is helpful for peer-to-peer provisioning across domains. At the top level, our architecture appears flat while the hierarchical structure is associated with large ISPs or ASs. We have also developed a distributed controller that attempts to maximize the effective throughput seen by the entire system and adapts to fluctuating load patterns. The CH-nodes close to the host networks are responsible for performing admission control. The edge routers maintain only aggregated state information about the flows and the core routers are completely stateless. The CH-nodes keep track of the intra- and inter-domain traffic patterns, and adapt aggregate reservations dynamically based on "Gaussian traffic predictors". The CH architecture can inter-operate with MPLS, OSPF and other queuing mechanisms like Core-Stateless Fair Queuing (CSFQ). The CH can be used to provision an ISP for VPN or VoIP traffic, and to achieve better QoS assurance across multiple domains.

Initial simulation results (slide #7) show that a single-node CH employing FIFO scheduling for the vBNS backbone topology has a saturation point of 1800 reservation requests/s for traffic load with Poisson arrival. Using aggregate scheduling, the throughput increases to 3500 reservation requests/s and the mean response time reduces by a factor of 4. Our simulations based on actual audio traces show that dynamic reservations based on a one-minute Gaussian predictor can achieve average packet loss of less than 1%, while incurring only 7% over-provisioning.