Next: Hot-potato versus Cold-potato routing
Up: Impact of multiple ISPs
Previous: Impact of multiple ISPs
Distribution of ISP path lengths
In this section, we further examine the distribution of the end-to-end
linearized distance that is accounted for by individual ISPs. We wish
to understand how the effort of carrying traffic end-to-end over a
wide-area path is apportioned between different ISPs. For each of the 13 nationwide ISPs in the U.S. listed in Section 3.4.1, we consider the set of paths that traverse one or more nodes in that ISP's network. For each such path, we compute the fraction of the end-to-end path that
lies within the ISP's network.
Figure 13:
CDF of the fraction of the end-to-end path that lies within individual ISP networks.
|
Figure 13 plots the CDF of this fraction for a few
ISPs. In each case, we consider the paths from the U.S. university
sources to the LibWeb data set. We observe that the distributions
look very different. For instance, the median fraction of the
end-to-end path that lies within Sprintlink is only about 0.35
whereas the corresponding fraction for UUNet is 0.75 and for
Internet2 is over 0.9. Internet2 is a high-speed backbone network
that connects many university campuses in the U.S. An end-to-end path
that traverses Internet2 typically originates and terminates at
university campuses. Therefore, the Internet2 backbone accounts for
an overwhelming fraction of such end-to-end paths. UUNET accounts for
a larger fraction of the paths that traverse its backbone than any
other commercial ISP we considered. This may reflect the close
relationship between UUNET's parent company, Worldcom (which runs the
vBNS backbone [29]), and academic sites.
The much smaller fraction in the case of Sprintlink is harder to
explain definitively. From our conversations with people at
Sprint [3,10], we have learned that academic
sites are not their major customers, so Sprintlink participates
minimally in carrying academic traffic. The location of our traceroute
sources at academic sites may explain why Sprintlink only accounts for
a small fraction of the end-to-end path.
We stress, however, that the point of our analysis is not to make
general claims about certain ISPs being better or worse than
others. Rather it is to show that geographic analysis of end-to-end
paths yields interesting insights into the role played by multiple
ISPs in specific contexts (e.g., academic sites) and that these
insights are consistent with our intuition.
Next: Hot-potato versus Cold-potato routing
Up: Impact of multiple ISPs
Previous: Impact of multiple ISPs
Lakshminarayanan Subramanian
2002-04-14