Academia.eduAcademia.edu

Outline

On the incompleteness of the AS-level graph

2012

https://doi.org/10.1145/2398776.2398803

Abstract

This paper has three parts: first, it presents an up to date characterization of the BGP visibility obtained from the three largest public sources: routeviews, RIPE, and PCH. Next, it proposes a criterion for a good monitoring infrastructure: one in which no AS is more than one provider-customer hop away from a monitor. It derives for the pubicly available dataset, the minimal set of ASes that meet this coverage goal. It discusses the properties of ASes that should be monitored. The question of BGP monitor placement is important, hard to study, and not often examined. This paper sheds important light on this question. Weaknesses: The paper's results are somewhat misstated, and the ultimate output is an approximation of unknown quality. The first part of the paper is helpful but doesn't really contain anything new. The discussion about the differences between routeviews/RIPE and PCH is helpful for folks who are using that data. As regards Section 3, the results are presented in a strange way. The authors are solving a set-cover problem. Clearly there is no efficient algorithm to solve this problem (unless P = NP). So it is strange for the authors to discuss an "algorithm" for this problem. Rather, what they have done is apply brute-force search to one specific dataset and found that they can obtain the optimal solution for that dataset. In particular, the authors provide no guarantees that their algorithm will run in acceptable time on any other dataset. So the claims (eg in the abstract) that they have "designed a novel algorithm for selecting the optimal number of ASes" should be removed. Further, the results are optimal only with respect to the constraint that no AS is more than one producer-consumer hop away from a monitor. This provides no guarantees on the fraction of links that will be observed using this monitor set. It is simply a heuristic that may or may not work well. The authors have not considered the effects of missing p2c links in their dataset. It would seem that the optimal set remains a set cover when missing links are added (good) but is no longer optimal (bad). The authors should acknowledge this fact and discuss the limitations due to missing p2c links.