From: Chandrika Jayant (cjayant@cs.washington.edu)
Date: Tue Nov 02 2004 - 18:58:58 PST
“Quantifying the Causes of Path Inflation”
Written by
Reviewed by Chandrika Jayant
This paper looks for the major causes of internet path inflation by studying the effects of topology and routing policy at three different networking layers- intra-domain (parts of the network belonging to a single ISP), peering between neighboring ISPs, and inter-domain (passing through multiple intermediate ISPs). The authors conclude that the two biggest factors contributing to internet path inflation are inter-domain and peering policy.
The data collection and topology extraction are explained quite clearly, but I am still not convinced that the number and type of ISPs chosen are representative of the Internet. Maybe they should have collected trace data from more than three days to produce better averages and weed out more bad results. The authors claim that the subset of ISPs chosen is representative because they include most large providers in the network. Some more discussion would be useful here. At first I thought that using DNS for geolocation would not be adequate, but the authors seem to take care of most possible incorrect data by looking for discrepancies. Other geolocation schemes could be used in the future to better approximate the delay-distance relationship.
Optimal versus observed values of inflation are compared in each of the 6 presented possible causes. I was not clear on how well isolation between the controls and the variables was maintained.
My main issue with the paper was the absence of traffic data. This prevents extending the granularity from paths to packets which experience inflation. Practically, this is what really matters in terms of network utilization and congestion problems- I’d like to know what is really causing what.
One impressive result of the paper is that ISP’s appear willing to cooperate to achieve low latency routing. Internet path inflation can be greatly reduced if ISP’s are aided in the engineering of better peering points and AS-path selection. Could other distance metrics besides shortest AS-path be used? BGP, Border Gateway Protocol, the inter-domain routing protocol in 2003, makes it difficult to implement sound global routing decisions. In peering and inter-domain routing, more global routing information is needed for robustness (suggested: geographical data). The obstacles here are technical and not commercial, which is helpful to distinguish because different inter-domain routing policies like no-valley and prefer-customer seem to make little difference.
This archive was generated by hypermail 2.1.6 : Tue Nov 02 2004 - 18:59:10 PST