Analysing the Effects of Routing Centralization on BGP Convergence Time

TL;DR

This paper presents an analytical model to evaluate how routing centralization via SDN influences BGP convergence time, providing insights into performance improvements in inter-domain networks.

Contribution

It introduces the first Markovian model for inter-domain SDN routing, analytically quantifying BGP convergence time under various network configurations.

Findings

Routing centralization can significantly reduce BGP convergence time.

Analytic results vary with SDN penetration, topology, and BGP settings.

Model offers scalable performance evaluation beyond simulations.

Abstract

Software-defined networking (SDN) has improved the routing functionality in networks like data centers or WANs. Recently, several studies proposed to apply the SDN principles in the Internet's inter-domain routing as well. This could offer new routing opportunities and improve the performance of BGP, which can take minutes to converge to routing changes. Previous works have demonstrated that centralization can benefit the functionality of BGP, and improve its slow convergence that causes severe packet losses and performance degradation. However, due to (a) the fact that previous works mainly focus on system design aspects, and (b) the lack of real deployments, it is not clearly understood yet to what extent inter-domain SDN can improve performance. To this end, in this work, we make the first effort towards analytically studying the effects of routing centralization on the performance of inter-domain routing, and, in particular, the convergence time of BGP. Specifically, we propose a Markovian model for inter-domain networks, where a subset of nodes (domains) coordinate to centralize their inter-domain routing. We then derive analytic results that quantify the BGP convergence time under various network settings (like, SDN penetration, topology, BGP configuration, etc.). Our analysis and results facilitate the performance evaluation of inter-domain SDN networks, which have been studied (till now) only through simulations/emulations that are known to suffer from high time/resource requirements and limited scalability.