Check out the new USENIX Web site. next up previous
Next: Multiple sources in the Up: Effect of network location Previous: Effect of network location


Paths from a single source

We consider paths from our traceroute sources in U.S. universities to two varied set of end-hosts: UnivHosts and TVHosts. Many of the hosts in UnivHosts (including our sources) connect to the Internet2 high-speed backbone via a local GigaPOP. So much of the wide-area path between our sources and a host in UnivHosts traverses the Internet2 backbone. On the other hand, TVHosts is a more diverse set that includes hosts located in various commercial networks (AOL, MSN, @Home, etc.) as well as university campuses. So the wide-area paths from our sources to the hosts in TVHosts typically traverse one or more commercial ISP backbones.

Figure 3: CDF of distance ratio for paths from UC Berkeley to UnivHosts and TVHosts.
\begin{figure}
\centerline{\psfig{file=figs/berkunivcli.eps,height=3.0in,width=4.5in,angle=0}}
\vspace{-0.20in}
\textbf{}
\end{figure}

This difference between the two groups of destination hosts is reflected in the cumulative distribution function (CDF) of the distance ratio for the two cases. As Figure 3 shows (for source in UC Berkeley), the distance ratio is close to 1 for many of the destinations. The ratio is 1.1 or less (corresponding to a linearized distance that exceeds the end-to-end geographic distance by no more than 10%) for 55% of the destinations in UnivHosts and 45% in TVHosts. This finding is consistent with the rich Internet connectivity of the San Francisco Bay Area (where UC Berkeley is located). The area includes several public Internet exchanges (e.g., MAE-West, PAIX, etc.) as well as private peering points. So a path from the UC Berkeley host to a destination host is often (but not always) able to transition to the latter's ISP within the SF bay area itself. So there is little need to take a detour through another city just to transition to the destination's ISP. There is a far more pronounced difference between the UnivHosts and TVHosts cases if we look at the tail of the distribution. For instance, at the 90th percentile mark, the distance ratio is 1.41 in the case of UnivHosts but 1.72 in the case of TVHosts; in other words, the detour is 1.75 times as large for TVHosts destinations as it is for UnivHosts (72% versus 41%). The paths to some of the hosts in TVHosts tend to be more circuitous because they traverse multiple commercial ISPs whose peering relationships may cause detours in the end-to-end path. We discuss this issue in more detail in Section 5. We observe qualitatively the same trends for other university sources as well; i.e., the distance ratio tends to be smaller for paths leading to UnivHosts compared to TVHosts.
next up previous
Next: Multiple sources in the Up: Effect of network location Previous: Effect of network location
Lakshminarayanan Subramanian 2002-04-14