Multiple sources in different locations

We consider paths from sources in three geographically distributed locations in the U.S.: Stanford, Washington University at St. Louis (WUSTL), and the University of North Carolina (UNC). The destination set is LibWeb, which is a larger and more diverse set than the UnivHosts set considered in Section 4.1.2.

Figure 5: CDF of distance ratio for paths from multiple sources to LibWeb.

As shown in Figure 5, the distance ratio tends to be the smallest for paths originating from Stanford and the largest for those originating from WUSTL. Stanford, like Berkeley, is located in the San Francisco Bay area, which is well served by many of the large ISPs with nationwide backbones. In contrast, WUSTL is much less well connected. Almost all paths from WUSTL enter Verio's network in St. Louis and then take a detour either to Chicago in the north or Dallas in the south. At one of these cities, the path transitions to another major ISP such as AT&T, Cable & Wireless, etc. and proceeds to the destination. Any detour is particularly expensive in terms of the distance ratio because the central location of St. Louis in the U.S. means that the geographic distance to various destinations is relatively small. In general, paths (such as those from WUSTL) that traverse significant distances in the backbones of two or more large ISPs tend to be more circuitous than paths (such as those from Stanford) that traverse much of the end-to-end distance in the backbone of a single ISP (regardless of who the ISP is). One example of a highly circuitous path we found involved two large ISPs, Verio and AT&T. The path originates in WUSTL in St. Louis and terminates at a host in Indiana University, 328 km away. However, the geographic path goes from St. Louis to New York via Chicago, all on Verio's network. In New York, it transitions to AT&T's network and then retraces its path back through Chicago to St. Louis, before finally heading to Indiana. The linearized distance is 3500 km, more than 10 times as much as the geographic distance. We examine the impact of multiple ISPs in greater detail in Section 5. While the specific findings pertaining to Stanford and WUSTL may not be important in general, our results suggest that the distribution of the distance ratio is consistent with our intuition about the richness of connectivity of hosts in different geographic locations.