Distilling Router Data Analysis for Faster and Simpler Dynamic IP Lookup Algorithms

We consider the problem of fast IP address lookup in the forwarding engines of Internet routers. Many hardware and software solutions available in the literature solve a more general problem on strings, the longest prefix match. These solutions are then specialized on real IPv4/IPv6 addresses to work well on the specific IP lookup problem. We propose to go the other way around. We first analyze over 2400 public snapshots of routing tables collected over five years, discovering what we call the "middle-class effect" of those routes. We then exploit this effect for tailoring a simple solution to the IP lookup scheme, taking advantage of the skewed distribution of Internet addresses in routing tables. Our algorithmic solution is easy to implement in hardware or software as it is tantamount to performing an indirect memory access. Its performance can be bounded tightly in the worst case and has very low memory dependence (e.g., just one memory access to off-chip memory in the hardware implementation). It can quickly handle route announcements and withdrawals on the fly, with a small cost which scales well with the number of routes. Concurrent access is permitted during these updates. Our ideas may be helpful for attaining state-of-art link speed and may contribute to setting up a general framework for designing lookup methods by data analysis.