From: Ethan Katz-Bassett (ethan@cs.washington.edu)
Date: Tue Nov 02 2004 - 18:12:18 PST
In this article, the authors present their work in characterizing the causes
of path inflation. They start from the observation that many paths in the
Internet are much longer than necessary. They then apply various techniques
to traces to attempt to break the inflation down into its component causes.
ISPs make this characterization difficult by having little incentive to
reveal their policies. The authors view path inflation as the aggregate of
inflation caused by topology and routing policy at three levels-within an
ISP, between two neighboring ISPs ("peering"), and across the sequence of
ISPs in a route.
Their inference techniques appear quite useful. For instance, they discover
that some ISPs use a load-balancing peering policy.
They find that routing at the peering and inter-ISP level, and topology at
the inter-ISP level cause most of the inflation. Interestingly, they did
not find that the economic incentives of competing ISPs play a large role
and actually have results showing lots of cooperation. I am interested in
the role of economic factors and look forward to the results from the groups
working on the SimISP project. Rather than economics, they conclude that
limitations in BGP keep ISPs from making well-informed routing decisions.
The work was trace-based, and so I am curious how consistent the results
would be across other data sets/ how representative their data was. One
curious aspect of their ISP set is that the two smallest are both in Oregon.
I wonder about the accuracy of their technique of mapping to a city-level
based on router naming conventions. Their techniques force them to exclude
ISPs that use MPLS, which is widespread.
The paper argues for the need for BGP to aid ISPs in path selection.
Interestingly, the authors suggest adding explicit geographic data to route
advertisements, allowing ISPs to take advantage of the strong connection
between geography and delay.
This archive was generated by hypermail 2.1.6 : Tue Nov 02 2004 - 18:12:24 PST