Inter-domain routing under scrutiny: routing models and alternative routing architectures

Much of the mystery behind inter-domain routing in the Internet comes from its complexity: A large number of independently administered autonomous systems (ASs), interactions between intraand inter-domain routing protocols, manifold routing policies, diverse peering structures between ASs, etc. The thesis at hand aims to improve the understanding of inter-domain routing in general, presents novel approaches for the construction of inter-domain routing models with predictive capabilities, and suggests an alternative routing architecture for a future Internet. We start with a comprehensive study of how sensitive routing optimality and diversity metrics are to factors such as policies, AS size, topology, and IGP weights, etc. Our findings reveal that intradomain factors only have marginal impact on global path properties while routing policies and AS size (in number of routers) are the dominating factors. Moreover, it is apparently hard to improve the global properties of route selection by existing means, i.e., tweaking BGP attributes, changing iBGP graphs, etc. Based on the obtained insights, we then discuss how to construct models of inter-domain routing with predictive capabilities regarding BGP paths. We start by showing the importance of considering more than one router per AS and introduce quasi-routers to capture path diversity as seen in observed routing data. Relying on the abstraction of quasi-routers, we show that our model provides accurate predictions for unobserved routes. Regarding routing policies our work then reveals that the granularity of actual routing policies is close to per-neighbor although the widely used AS relationship model fails to provide consistency between the routes propagated in our routing model and those seen in observed data. Given several shortcomings of Internet routing such as routing table growth, high update rates, or lack of traffic engineering, we finally discuss long-term solutions for the future Internet. Our research in this area consists of two distinctive parts. First, we present Trellis, a network testbed that can be used by network researchers to implement, test, and evaluate (clean-slate) solutions for a future Internet. As such, it allows to partition a physical network into multiple logical or virtual networks, where each virtual network can define its own topology, routing protocols, and forwarding tables. Second, we introduce HAIR , a scalable routing architecture for a future Internet. HAIR uses a hybrid “edge-based” approach to reconcile networkand host-based routing architectures and is based on the following ideas: (i) use of hierarchical routing, (ii) separation of locators from identifiers, and (iii) use of a hierarchical mapping system. We estimate the expected benefits of deploying HAIR in today’s Internet and report our promising experiences with a proof-of-concept implementation of HAIR.