Limitations

We now discuss the limitations of our study arising both due to the inherent limitations of geographic information and due to limitations of our experimental methodology.

1. Geography does not determine performance: There is not a perfect relationship between geographic distance and network performance. It is possible that a circuitous path yields better performance than a less circuitous one. For instance, the most optimal path between certain countries may be via the U.S. even if that means a large detour in geographic terms. However, in Section 4.5, we show that there exits a strong correlation between the minimum end-to-end delay between two end-hosts and the linearized distance of their connecting path. In light of this, we view our geographic analysis of network paths as providing (a) hints on paths that are potentially anomalous and should be examined more closely to determine if they are indeed anomalous, (b) an indication of how much improvement there could be in end-to-end latency if a non-circuitous path between source and destination were feasible, and (c) a way to quantify network properties such as hot-potato routing, which may provide new insight into these properties.
2. IP-level topology is incomplete: Our linearized distance computation only considers the router-level (i.e., IP-level) topology. We have no way of discovering the underlying physical topology (which may be based on ATM, SONET, or other technologies), so in general we would underestimate the linearized distance. While this is a limitation of our methodology, we note that the trend in high-speed networks (OC-48 and faster) is away from separate layer-2 and layer-3 architectures (e.g., IP-over-ATM) and towards an all-IP network [15]. This trend increases the applicability of our methodology.