From: Masaharu Kobashi (mkbsh@cs.washington.edu)
Date: Tue Nov 02 2004 - 17:02:28 PST
1. Main result of the paper
The paper analyzes the causes of path inflation on the Internet
categorizing the potential causes into three stages: intra-domain,
peering between ISPs and inter-domain stages. It concludes that
the major causes of the path inflation are in the peering and
inter-domain stages. It also claims that the path inflation can
be significantly reduced by introducing good tools that enable
ISPs to engineer good peering points and make right AS-path
selection.
2. Strengths in this paper
The authors cleverly separates the potential causes into three stages
and each stage into two categories.
Although their findings are along the intuitive guesses by many people
(that the inflation takes place at the borders between ISPs), the paper
deserves credit for discovering the fact by the empirical evidences
based on the study of substantial portion of major ISPs.
3. Limitations and suggested improvements
The authors resort to the calculation of weights of links to find
the intra-domain policies. But their scheme of using static linear
inequations cannot capture dynamic load-balancing activities which
many ISPs deploy. The authors' finding of AT&T's strange behavior
(making seemingly unnecessary round trip from San Francisco to Seattle)
is one of the examples.
To find the intra-domain policies, it is best to directly ask managers
of some ISPs, although it does not give scientific looking to the paper.
If you cannot interview many of them, small number of samples will do,
since ISPs take similar behavior, each of them being bound by similar
economic and technical constraints. I do not think it is hard to find
those who can cooperate in revealing their policy to the extent which
is useful to the objective of the paper.
The authors concluded that the inter-domain routing decision is to
seek shortest AS-paths. But it can be quite possible that the shortest
AS-paths happen to be the least expensive paths and the actual
motivation of the ISPs is not to find the shortest AS-paths but to find
the least expensive paths. The latter can be the case since within a
single AS the announced costs to transit customers can be the same
regardless of peering point or the costs may not exactly correlated
to actual geographical distances/latencies of paths.
4. Relevance today and future
If in the future there are good tools for ISPs to find good paths,
as suggested by the authors, in terms of latency, there will still
be inflation problems if there are discrepancies between "good paths"
and the most economic paths.
Most ISPs will prefer minimum economic costs to the saving of around
10 milliseconds. It seems the major factors that determine
the inter-ISP relations are economic rather than the saving of
around 10 milliseconds in latency.
This archive was generated by hypermail 2.1.6 : Tue Nov 02 2004 - 17:02:29 PST