BGP Session Lifetime Modeling in Congested Networks Li Xiao † , Guanghui He † , and Klara Nahrstedt Department of Computer Science, 201 N. Goodwin University of Illinois at Urbana-Champaign, Urbana, IL 61801 Email: {lixiao, ghe, klara}@cs.uiuc.edu Abstract The reliable Border Gateway Protocol (BGP) is essential for supporting high quality Internet data communication. In present Internet, due to the lack of differentiation mechanism in packet forwarding, BGP sessions are sensitive to severe network congestion. It is thus important to understand the reliability of BGP in congested networks, for the purposes of system reliability evaluation and failure avoidance. In this paper, we investigate the lifetime of BGP sessions in two types of network congestion scenarios: (1) the TCP bandwidth saturation caused by traffic engineering failures; (2) the UDP bandwidth saturation caused by worm attacks. Using statistical analysis based on intensive simulation results, we find that in most cases the BGP session lifetime can be characterized using exponential distributions and Weibull distributions. In the case of TCP bandwidth saturation, if all TCP connections have the same round trip time, the tail of the BGP lifetime tends to be power-law. Furthermore, we propose an approximate model for the expected lifetime of BGP sessions, and show that by slightly changing the TCP retransmission parameters, the robustness of BGP sessions can be improved significantly. I. I NTRODUCTION Border Gateway Protocol (BGP) [1] is one of the most important network protocols for today’s Internet. The routers that run BGP and exchange routing information directly with each other are called BGP peers. Two BGP peers communicate with each other via a BGP session. The robustness and stability of BGP sessions are crucial in providing high-quality Internet data communication support. In this paper, we investigate the reliability of BGP sessions. A. BGP Failures in Congested Networks Multiple reasons can cause BGP session failure (or reset), such as network congestion, network failures, etc. In order to decrease the failure probability of BGP sessions or even prevent the failure from happening, we have to identify the conditions, in a quantitative way, under which BGP sessions could fail. Thus, it is important to understand the characteristics and the influential factors of BGP session failures in various stressful network scenarios. This work was supported by NSF under contract number NSF ANI 03-23434. Any opinions presented here do not necessarily reflect the views of the National Science Foundation. † At present, Li Xiao is with Google Inc, and Guanghui He is with Microsoft Corp. 1