No abstract

Border Gateway Protocol allows Autonomous Systems (ASs) to apply diverse routing policies for selecting routes and for propagating reachability information to other ASs. Although a significant number of studies have been focused on the Internet topology, little is known about what routing policies network operators employ to configure their networks. In this paper, we infer and characterize routing policies employed in the Internet. We find that routes learned from customers are preferred over those from peers and providers, and those from peers are typically preferred over those from providers. We present an algorithm for inferring and characterizing export policies. We show that ASs announce their prefixes to a selected subset of providers. The main reasons behind the selective announcement are the traffic engineering strategy for controlling incoming traffic. The impact of these routing policies might be significant. For example, many Tier-1 ASs reach their (direct or indirect) customers via their peers instead of customers. Furthermore, the selective announcement routing policies imply that there are much less available paths in the Internet than shown in the AS connectivity graph. We hope that our findings will caution network operators in choosing the selective announcement routing policy for traffic engineering. Finally, we study export policies to peers and find that ASs tend to announce all of their prefixes to other peers. To the best of our knowledge, this is the first study on systematically understanding routing policies applied in the Internet.

No abstract

This paper studies the contributions of routing dynamics to the duration of long-lived end-to-end Internet path failures. Studies have shown that end-to-end Internet failures (periods of prolonged packet loss) are widespread. These failures are typically attributed to either congestion or routing dynamics. Unfortunately, the extent to which congestion and routing dynamics contribute to long-lasting path failures, and the effect of routing dynamics on end-to-end performance, are not well understood. This paper uses a joint analysis of active measurements and routing data to characterize end-to-end failures observed over one month on a topologically diverse Internet testbed. We find that routing dynamics coincide with most prolonged end-to-end failures, suggesting that routing dynamics contribute significantly to the duration of these failures. We also find that most long-lived end-to-end path failures that coincide with routing dynamics are caused by BGP convergence or instability. Our results provide new insights into the effects of routing instability on end-to-end Internet path performance.

No abstract