Using forgetful routing to control BGP table size

Abstract

Running the Border Gateway Protocol (BGP), the Internet's interdomain routing protocol, consumes a large amount of memory. A BGP-speaking router typically stores one or more routes, each with multiple attributes, for more than 170,000 address blocks, and growing. When the router does not have enough memory to store a new route, it may crash or enter into other unspecified behavior, causing serious disruptions for the data traffic. In this paper, we propose a new mechanism for routers to handle memory limitations without modifying the underlying routing protocol and without negatively affecting convergence delay. Upon running out of memory, the router simply discards information about some alternate routes, and requests a "refresh" from its neighbors later if necessary. We present an optimal offline algorithm for deciding which alternate routes to evict, and explore the trade-off between memory size and refresh overhead using a large BGP message trace. Based on these promising results, we design and evaluate efficient online algorithms that achieve most of the performance benefits. We believe that our scheme can significantly improve the scalability and robustness of IP routers in the future.

Pathlet routing

“…The techniques of [16] may be very easy to apply in our setting to reduce control plane memory use from O(δ ) to O( ) per pathlet, where δ is the number of neighbors and is the mean path length. Routers could also pick dissemination paths based on heuristics to predict stability, which for common failure patterns can significantly reduce the number of update messages [12].…”

Offloading routing complexity to the Cloud(s)

“…To improve the amount of traffic and CPU usage during the PE process, there are many proposed schemes [22], [23], which allow a peer to dynamically express its interest on advertisements of specific prefixes instead of full-table exchange without introducing significant convergence delays. These mechanisms are particularly useful for enterprise network scenarios where border gateways may be interested in routes for a few prefixes due to their policies while accepting default route from its provider for the rest.…”

Scalable routing on flat names

“…This may be acceptable for low-degree graphs or if routers with higher degree are more well-provisioned. We can reduce control state to Θ( √ n log n) by either forgetting the unused announcements as in Forgetful Routing [24], or by preventing them from being sent by having each node inform its neighbors of the "radius" of its vicinity.…”