From: Michael J Cafarella (mjc@cs.washington.edu)
Date: Wed Nov 03 2004 - 03:44:42 PST
Quantifying the Causes of Path Inflation
By Neil Spring, Ratul Mahajan, and Thomas Anderson
Review by Michael Cafarella
CSE561
November 3, 2004
Main result:
It's a well-known result that end-to-end internet paths are much longer than
they have to be. The authors use an interesting instrumentation mechanism to
figure out why this might be. They use a series of vantage points on the internet
to contact a set of well-known targets, and record all the paths that the packets
take. Each hop in a route's path involves a machine, which is owned by an ISP.
They authors then gather BGP information from many vantage points to figure out
the ISP connectivity graph.
Using all this information, the authors can deduce the internal routing policies
of ISPs. They then figure out where in the hop structure most of the extra
time is being spent, and how those policies are contributing.
Strengths of paper:
The system they use to gather the data is very impressive. There are no public
sources of this stuff, and they seem to gather an accurate map all by themselves.
They can use the data to find some interesting conclusions (like the
possibly-misconfigured link to Seattle).
Limitations, other problems:
It's not clear that early-exit, late-exit, etc, are sufficient to model an
ISP's motivations. A more interesting model might let us say something more
powerful about what the ISP is doing.
Further, while the apparatus is very impressive, it's also unwieldy. Better
if we didn't need it, or could make due with fewer traceroutes.
Possible improvements:
Reduce the data needed, develop more sophisticated ISP policy models.
Modern relevance, future work:
Routing in the face of real-world barriers, such as ISP boundaries, was
not considered by the original Internet design. Since many networking ideas
fail when it comes to deployment, a better understanding of ISPs and real-world
deployment issues is probably helpful.
This archive was generated by hypermail 2.1.6 : Wed Nov 03 2004 - 03:44:42 PST