Dynamics of RIB Usage at An Edge Router

Previous studies [2], [4], [7], [8] have shown there is a small set of popular prefixes. These prefixes carry a high volume of traffic, experience few BGP route changes, and have other desirable characteristics. However, these popular prefixes constitute only a tiny fraction of the BGP routing table and the behavior of the remaining prefixes has received considerably less attention. For example, do the other prefixes oscillate between temporarily popular and then inactive? If one considers the top X prefixes ranked by traffic volume, will the set be fixed or vary wildly? The answers to these questions have important implications for route scalability and prefix aggregation, determine whether proposed prefix caching and/or FIB caching techniques would be deployable, and could help guide BGP design changes as well as future routing protocol designs. According to a recent IAB report [5], Internet routing, specifically the FIB (Forwarding Information Base), faces scalability issues due to the ever expanding size of the global routing table. This size increase is fueled, in part, by the injection of prefixes into the routing table due to multihoming and the need for traffic engineering. Large FIBs affect lookup speed and the price of routers, which now become more expensive due to the need for increased memory. Moreover, this increase threatens to push existing routers over their current limits, pushing ISPs into forced upgrades. These large tables consist of both the popular prefixes whose characteristics have been studied and the vast set of other prefixes for which much less is known. Several approaches have been suggested to overcome the routing scalability problem [1], [3]. Many of these approaches call for routers to route packets while storing less than the full FIB. With reduced FIBs it is important to carefully choose the prefixes that are stored in the routers FIB. Approaches such as LISP [3], Map-n-encap, Viaggre [1] aim to solve the scalability problem by storing reduced FIBs at the routers. This works well for the small set of popular prefixes, but are those prefixes sufficient? Choosing the appropriate prefixes to store at the routers requires an in-depth, fine-grained understanding of the prefix behavior. In this work, we study the dynamics of prefix usage on provider links between a regional ISP and two tier-1 ISPs. Using packet captures from a regional ISP, we associate packets with the BGP prefix used to forward them. We define a new metric, the mean rank difference, to capture how busy a prefix is. With this metric, we look at how a prefix maintains its rank throughout the day. As expected, the most dominant (popular) prefixes during the day tend to maintain their rank but the real objective is better understand the majority of prefixes, not just the most popular ones.