Geographic fault tolerance of ISPs

An important component of studying Internet routing is to understand its fault tolerance aspects. Fault tolerance of a network is normally studied at the granularity of router or link failures. However such a failure model does not capture the fact that two seemingly independent routers can be susceptible to correlated failures. We ask the question: what is the tolerance of an ISP's network to a total network failure in a geographic region, i.e., a failure that affects all paths traversing the region? We refer to such a failure as a geographic failure. Potential reasons for such a failure include natural calamities such as earthquakes or power blackouts. By using the geographic location information of the routers, we can identify routers that are co-located and thereby construct a geographic topology of an ISP. In this topology, each geographic region is associated with a node and an edge between two nodes signifies the existence of at least one long-haul backbone link that connects the corresponding geographic regions. We obtained the geographic topologies for 9 of the 13 major ISPs listed in Section 3.4.1 from the CAIDA MapNet site [24]. These are: AT&T, Cable and Wireless, Sprintlink, Genuity, Qwest, PSINet, UUNet, Verio and Exodus. Many of these topologies are obtained from information published at the ISPs' Web sites and are between 6-12 months out of date. Although it may be possible to construct an ISP's geographic topology using extensive traceroute measurements, it would be hard to assess the completeness of the constructed topology. Hence we restrict ourselves to the geographic topologies obtained from CAIDA. However, as acknowledged by CAIDA [24], it is possible that these topologies may themselves be incomplete. This may be due to limited tracing or the presence of backup paths in routing. We will perform our analysis under the assumption that these topologies are reasonably complete and only have a few missing links.