The End-to-End Effects of Internet Path Selection Review

The End-to-End Effects of Internet Path Selection Review
Review By: Troy Ronda

There are many possible paths, at a given time, between two arbitrary
hosts on the Internet. The quality of these paths depend on many
factors, limiting their end-to-end performance. If packets become lost
or delayed on the path, throughput will decrease. Both economic and
technical reasons lead to the conclusion that Internet routing is
non-optimal. Some providers will refuse to carry traffic without a
contract, for example. This paper studies the possibility of alternate
paths. Five datasets were collected containing a large number of paths.
The authors collect measurement for each path including round-trip time,
loss rate, and bandwidth. Each measured path is compared to synthetic
alternate paths – built by composing measurements from multiple
connected paths. These alternate paths potentially use different
providers and may have different degrees of connectivity. By composing
many paths between hosts, a hypothetical path that routes around
inefficiency can be found. The authors examine round-trip latency, drop
rate, and in some cases bandwidth on these hypothetical paths. These are
compared to the default path provided by Internet routing. It is found
that 30-70% of hypothetical paths are higher quality. Two hypotheses are
are presented for this finding. First, superior alternate paths are
caused by avoiding parts of the Internet of bad quality. Second,
superior alternate paths are caused by avoiding congestion. The authors
show the former is false and the latter is true.

This paper is a harder read than the previous ones from Savage et al. I
am not sure if the material was dry, the font size too small, or that
there is too much content squeezed into the paper. I did find the
results, on one hand, not surprising, as Internet routing is not
optimal. On the other hand, the authors are routing between individual
hosts meaning the aggregate last mile cost must be high. It is amazing
that even 30-70% of these paths are better than the default path
provided by Internet routing. This paper is good support for using
overlay networks as a routing strategy (on the assumption that
congestion will continue to be a problem in the Internet). I also found
the finding interesting, that the quality difference between alternate
paths and the default path is higher during the North American
east-coast work day. This further adds to the conclusion that alternate
paths are avoiding congested networks. I find this paper good evidence
that the AS routing decisions are not optimal for routing, though they
might be optimal for economic reasons (I doubt they are even optimal for
that).

This paper gives no advice on how to find these alternate paths online.
The second problem is that the authors stated they wanted to answer the
question of “How good is Internet routing from a user’s perspective.” I
argue that this question is not entirely answered. It is true that they
have shown that the situation is non-optimal but how much does this
affect the average end-user of the Internet. (i.e. Do people not in the
field care?) I cannot find the answer in this paper, normally I wouldn’t
say anything but this time the authors made a point of declaring this
their goal. Going back to overlay networks from the previous paragraph.
What would happen if we switched the load to one of these alternate
paths? Would the path quality decrease because it now has a higher load?
Is this one of the factors involved? I do not think the authors made
this situation clear enough.
Received on Thu Nov 03 2005 - 09:19:47 EST