Interdomain routing as social choice (2006)
| Venue: | IN PROCEEDINGS OF INCENTIVE-BASED COMPUTING (IBC |
| Citations: | 3 - 0 self |
BibTeX
@INPROCEEDINGS{Salihoglu06interdomainrouting,
author = {Semih Salihoglu},
title = {Interdomain routing as social choice},
booktitle = {IN PROCEEDINGS OF INCENTIVE-BASED COMPUTING (IBC},
year = {2006},
publisher = {}
}
OpenURL
Abstract
Interdomain routing is an essential component of the Internet, which deals with connect-ing two machines in the Internet through a set of links. Today this is done through BGP, the de facto protocol of the Internet. Previous studies focus on the stability of BGP type proto-cols. The current thesis focuses on interdomain routing from a social choice perspective and is divided into two main parts. In the first part we analyze interdomain routing considering optimality. We view interdomain routing as a system, defining a social choice rule that ag-gregates individual preferences of autonomous systems (ASes) to select a set of routing trees. An interdomain routing protocol, then, is a mechanism implementing a social choice rule. We study the desirable properties of such social choice rules in the interdomain routing context. By presenting major incompatibilities between certain properties and studying their imple-mentability limitations, we reveal fundamental tradeoffs that must be made when designing a new interdomain routing protocol. We also define BGP’s implicit social choice rule and outline its properties. To overcome the limitations of interdomain routing outlined in our social choice analysis, we investigate ways of designing protocols that do not output routing trees. We first propose BGPk, a family of BGP type protocols, that relax the current BGP convention of se-lecting the best available route to selecting the highest k available routes. We show that even 1 though the output of BGPk is not necessarily a routing tree, it fails to solve BGP’s oscillation problem. We then propose HiRANP (Highest Ranked Neighbor Protocol) as a more effective alternative and show that it is scalable, efficient, and convergent.







