Hot-potato versus Cold-potato routing

Finally, we investigate whether geographic information can be helpful in assessing whether ISP routing policies in the Internet conform to either hot-potato routing or cold-potato routing. In hot-potato routing, an ISP hands off traffic to a downstream ISP as quickly as it can. Cold-potato routing is the opposite of hot-potato routing where an ISP carries traffic as far as possible on its own network before handing it off to a downstream ISP. These two policies reflect different priorities for the ISP. In the hot-potato case, the goal is to get rid of traffic as soon as possible so as to minimize the amount of work that the ISP's network needs to do. In the cold-potato case, the goal is carry traffic on the ISP's network to the extent possible so as to maximize the control that the ISP has on the end-to-end quality of service. In general, an ISP's routing policy would lie somewhere in between the extremes of hot-potato and cold-potato routing.

Figure 14: CDF of the fraction of the end-to-end path that lies within the first and second ISP networks in sequence.

We consider the set of paths from U.S. sources to TVHosts. For each path that traverses two or more major ISPs (with nationwide backbones), we compute the fraction of the end-to-end path that lies within the first major ISP (ISP1) and the second major ISP (ISP2) in sequence. We use these fractions as measures of the amount of work that these ISPs do in conveying packets end-to-end. The distributions of these fractions is plotted in Figure 14. We observe that the fraction of the path that lies within the first ISP tends to be significantly smaller than that within the second ISP. For instance, the median is 0.22 for the first ISP and 0.64 for the second ISP. This is consistent with hot-potato routing behavior because the first ISP tends to hand off traffic quickly to the second ISP who carries it for a much greater distance. Figure 14 also plots the distributions of the path lengths in the case where the first ISP is Sprintlink. We find that the difference between the ISP1 and ISP2 curves is even greater in this case. Again, this is consistent with hot-potato routing behavior on the part of Sprintlink for routes from academic locations.