Browse Theses

Resource reservation protocols were originally designed to signal end hosts and network routers to provide quality of service to individual real-time flows. More recently, Internet Service Providers (ISPs) have been using the same signaling mechanisms to set up provider-level Virtual Private Networks (VPNs) in the form of MPLS Label Switched Path (LSP). It is likely that the need for reservation signaling protocols will increase, and these protocols will eventually become an indispensable part of Internet service. Therefore, reservation signaling must scale well with the rapidly growing size of the Internet.

Over the years, there have been debates over whether or not there is a need for resource reservation. Some people have been advocating over-provisioning as the means to solve link congestion and end-to-end delay problems. The over-provisioning argument is largely driven by the expectation that the bandwidth price will drop drastically. From our investigation, however, we found that many end users have not been benefiting from over-provisioning: the current Internet has bandwidth bottleneck links that can cause long-lasting congestion and delay. At the same time, leased line cost has not been reduced sufficiently in a timely manner for many network providers to deploy high-speed links everywhere in their networks.

Applying resource reservation brings many benefits to the network users. Unfortunately, the current resource reservation framework has scalability problems in terms of storage, bandwidth, message processing and manageability. To address these problems, we first evaluate methods that are designed to improve the scaling properties in RSVP. Though some of the methods can reduce protocol processing overhead substantially, they do not reduce the total number of reservations in the network. Thus, we argue that merely enhancing the existing signaling protocols may not be sufficient.

Generally, scalability problem can be solved by building a hierarchy. Resource reservation signaling is no exception. Depending on traffic behavior and service requirements, we propose a hierarchical reservation model that will support reservation signaling capability at end-user's application layer as well as at network provider's backbone level. In the model, end users may use lightweight signaling protocols to setup reservations for short-lived real-time applications. Within the network, service providers, based on bilateral agreements, establish long-lasting and more static reservation “trunks” among each other. At the network edge or border, end-user reservations are aggregated into provider's reservation “trunks”, depending on user's qualification and network resource availability.

To explore our understanding on lightweight signaling, we introduce YESSIR, a simplified application-layer reservation protocol. It is designed to establish reservations for real-time streaming traffic. To simplify the processing at routers, YESSIR uses one-pass signaling sequence and allows data senders to initiate reservations. YESSIR also uses partial reservation and reservation retry techniques to speed up the setup. Our implementation results show that with proper protocol design and implementation, network routers can support a large number of user reservations (10,000 reservation requests per second on a FreeBSD prototype).

One of the most challenging aspects on provider-level signaling is that the protocol needs to be applicable and scalable to potentially all network providers in the Internet. After evaluating traffic traces from the Internet backbone, we derive a sink-tree algorithm, where the reservations from other providers following inter-domain routing path to a destination provider's network form a tree, rooted at the destination provider's border router. The sink-tree approach has the property that the maximum number of reservations at network routers is always O ( N ), where N is the total number of routing domains in the network. This should reduce the total number of inter-domain reservations to a manageable level. We present an inter-domain reservation protocol, BGRP, that is based on the sink-tree algorithm. BGRP also has several built-in features that allow fast setup and make it resilient to route flapping.