From: Charles Reis (creis@cs.washington.edu)
Date: Wed Nov 03 2004 - 07:43:14 PST
Quantifying the Causes of Path Inflation
Spring, Mahajan, Anderson, 2003.
The paper investigates possible reasons why Internet paths are often much longer than necessary, which is a difficult question given that ISPs do not openly share information about their topology and routing policies. Instead, the authors use a large set of measurements from diverse locations to infer how ISPs are configured, building models from this data to compare hypothetical ideal paths and actual paths. This is performed at three different granularities (intra-domain, peering, and inter-domain) to learn about the effects of ISP topology and policy at several different levels.
In addition to identifying that a large portion of path inflation occurs because of inter-domain routing policy, the techniques reveal or confirm many interesting aspects of ISP configurations, such as the use of early or late exit routing and no-valley and prefer-customer agreements between ISPs. By comparing the effects of each of these policies, the authors are able to conclude that the bulk of the problem comes from limitations of BGP for improving path selection. Given that other observations show that ISPs are willing to collaborate rather than purely compete, this strongly suggests new research in specifying BGP policies.
It is worth noting that the paper does a very good job building on existing work on measurement and inference and impact of routing policies. It is perhaps unclear whether computing ideal latencies based on geography is realistic, but it at least provides a good baseline for comparison. Finally, it is hard to judge the completeness or accuracy of the results, given the indirect measurements and limited views into ISPs. In this sense, confirmation from ISPs would be very useful.
This archive was generated by hypermail 2.1.6 : Wed Nov 03 2004 - 07:43:15 PST