From: Seth Cooper (scooper@cs.washington.edu)
Date: Wed Nov 03 2004 - 00:34:28 PST
The purpose of this paper is to attempt to figure out where the causes
of path inflation lie. Path inflation means that end-to-end paths
become longer than necessary in the network. The paper collects a large
amount of data to determine network topology and ISP peering. Then it
attempts to determine the the policies used by each ISP. Finally, it
attempts to discover how topology and policy affect path inflation at
the intra-domain, peering, and inter-domain levels. The paper concludes
that ISPs are trying to cooperate to give paths low latency, but
limitations of the Border Gateway Patrol prevent them from coordinating
their efforts and determining the best paths.
One strength of the paper is the amount of data gathered. It appears
that a lot of effort went into gathering data and making sure that it is
valid. The data was gathered over a period of three days, allowing
transient paths to be filtered from the dataset. Decisions such as
excluding data that may lead to underestimation of path inflation, such
as from ISPs that use virtual circuit technologies, give the data more
robustness.
A weakness of the paper is that there were a lot of ISP pairings for
which only one peering was seen. This makes it difficult to evaluate
the effect that the peering policy between the two connected ISPs would
have on path inflation. Although there was already a lot of data
gathered for thus study, it would be an improvement if enough data could
be gathered to better evaluate peering policies between all pairs of
connected ISPs.
One interesting result of this paper is that the table in Figure 9 is
not symmetric. This implies that some ISPs are treating traffic they
send to another ISP differently than they treat the traffic that ISP is
sending to them.
This paper is relevant because the paths chosen by routers have an
effect on the network as a whole. Paths shape the traffic and have an
impact on user experience. Whenever the network is choosing paths that
are unexpected, it is important to figure out why that is, and if it
needs to be corrected.
This archive was generated by hypermail 2.1.6 : Wed Nov 03 2004 - 00:34:29 PST