From: pravin bhat (pbhatgrad@hotmail.com)
Date: Wed Nov 03 2004 - 07:57:16 PST
Paper summary: The paper provides a statistical analysis of path inflation
in the internet today. Contributing factors are analysized at three levels
in the routing heirarchy - intra-domain routing, routing between adjacent
ISPs and inter-domain routing.
Paper Strengths:
The paper is impressive in its sheer ambition to analyze path inflation
in the internet without any co-operation from the ISPs. The authors do a
great job in leveraging previous work in the field and some new ideas in
reconstructing the network topologies, estimating routing policies, guessing
the link costs used in the shortest path calculations, cleaning and
simplifying the
noisy data, etc. All this using only traceroute data collected on end
systems.
The paper presents an through analysis of path inflation seen in the data
across various levels in the routing hierarchy and helps the reader
realize exactly what is causing the most damage.
The conclusions drawn from the data are surprisingly unintuitive - for
example one would guess that the path inflation caused by inter-domain
politics
is bound to be a lot more damaging that then the BGP policy to route packets
by
shortest hops across ISPs. Its also surprising to see that routing policies
in ISPs are not purely motivated by their financial interests.
Limitations and room for improvement:
I failed to understand how the authors could approximate expected/optimal
latency
as a linear function of geometric distance. At the very least one would
expect the
link speeds to vary across networks even if we were to assume the networks
were
experiencing uniform traffic loads. Did the authors assume that all networks
used the same link technology? Was the data pruned to remove outliers in
this
category - i.e. all satellite links?
Unfortunately analysis of a system as complex and heterogeneous as the
internet
involves factoring for innumerable special cases and pruning the data to
remove outliers. For example, the authors show incredible foresight in
detecting
and excluding MPLS networks from their data. Had the authors not accounted
for
this special case I would guess that most readers would have would failed to
spot it themselves. I have to wonder how many such complex special cases are
unaccounted for in the data collected for this experiment.
On a similar note it would have been nice to know if the authors processed
the
data using any statistical tools to search for abnormalities.
For example - I would like to know how the authors came across the subject
of their case-study - the link between AT&T and Sprint. It would be a lot
more
comforting to know that this case was spotted due to some statistical tool
as
opposed to the off-chance that one of the authors just happened to look at
the
right section of the data.
Future work and relevance:
This work is extremely relevant for the future and long overdue. Statistical
analysis of the underlying dynamics of the Internet is an extremely hard
problem. The complexity of the Internet has grown to the point where we have
to
move away from our existing model of relying on a few key
researchers/industry-experts
to guide future improvements based simply on their intuition.
The conclusions presented in the paper are unintuitive and probably even
surprised
the authors. Research along this direction will help future researchers to
spot
the key problems in the field and better guide their efforts towards solving
the more pressing issues of the day.
For future work I would like to see the experiment repeated on random three
day
periods. It would be interesting to see if the results presented in the
paper
can consistently reproduced.
I got the feeling that the authors were surprised to realize that there was
significant co-operation between the ISPs. This reflects poorly on the
co-operation level between ISPs and academia and shows more room for future
improvement in this relationship. Its just bizarre that the researchers have
to go
through such lengths to form a reasonable of model of the inner working of
ISPs when they should be able to simply ask the ISPs.
_________________________________________________________________
FREE pop-up blocking with the new MSN Toolbar – get it now!
http://toolbar.msn.click-url.com/go/onm00200415ave/direct/01/
This archive was generated by hypermail 2.1.6 : Wed Nov 03 2004 - 07:58:24 PST