CS268 Reading Review

Internet Routing Instability

Vern Paxson
Review by Feng Zhou
2/16/2003

The problem: This paper addresses the same problem as [Pax96], with the compliment approach. Instead of measuring routing stability using end-to-end probing, the authors collected BGP traces from major Internet exchange points and analyzed them. So the problem addressed here is: whether BGP performs well and keeps inter-AS routing tables stable.

Key points:

  1. The basic conclusion of the paper is: the inter-AS routing tables are NOT stable at all and therefore there are serious problems with wide-area Internet routing. Nearly 99% of routing updates are pathological and does not reflect actual topological change. Although a lot of the redundant updates are duplicate ones, which means they do not do much harm to actual routing behavior of core Internet routers, they do pose a significant burden on routing servers in charge of processing these update messages. The paper reports that during peak hours of update messages, the volume of messgea is large enough to overwhelm some high-end routers.
  2. Regarding the causes of routing instability, several observations made by the authors are interesting. First, the number of top level visible (addressable) prefixes are directly related the frequency of routing updates, because the default-tree of backbone routers maintains routing information for all these prefixes. Therefore prefix aggregation (or supernetting) can alleviate the problem, if widely executed. However, commercial ISPs and end users are reluctant to do this. And the prevalence of multi-home prefixes technically prevents aggregation. This example raises an interesting question: given the fact that the Internet design relies much on altruism, how does the commerciallized Internet without altruism perform in various aspects and how can we change the Internet to cope with these problems.
  3. In comparison with Pax96, the approach used in this paper naturally reveals more accurate and comprehensive results. However, we should see that as the Internet backbone becomes more and more distributedly administrated, doing measurements in this way requires more and more collaborative between different parties. As ISPs increasingly interconnect with each other, instead of connecting to the same backbone, it is becoming harder and harder to get all statistics of all routing updates.