acm sigcomm

Signal is not information, and however hard we try our networked life is not 100% efficient. In some extreme cases, one might ever wonder if some of these online conversations are worth taking place. One reasurring thought is that, if the value of communication is subject to debate, its cost on the other hand can probably be assessed objectively. No matter how boring is your train neighbor’s cellphone conversation, you can probably infer how much he or she will eventually pay for it. This paper may (or not) surprise you as it asks the apparently simple question “what is the real cost of carrying a given traffic flow?” and it tells that, well, it's perhaps more complicated than we think. As several reviewers pointed out, this is perhaps even more complicated than what a 6 page paper can tell. There are many good reasons (much better than the one I gave above) to compute this cost, primarily for an operator to optimize key decisions like the establishment of peering links. This is where the paper focus and it establishes that it is likely that the cost can be greatly reduce by modifying routes for a small portion of the traffic. Where it differs from traditional traffic engineering is that it does not aim at a previously agreed performance goals, but it uses the same means to minimize the overall cost. Not surprisingly, the cost model turns out to be an essential piece: for substantially the same gain, the fraction of traffic to reroute varies from 10% (for linear cost) to 30% (when a cooperative game theory following Shapley’s fairness axioms are used). This result more generally establishes that cost models matter, and also that they can be useful. One clear merit of this work is to make all of us aware that perhaps our community should be engaging in understanding which cost model can and should be used. Major questions remaining to answer are (1) the impact of the congestion and feedback loop, (2) the roles that content providers and CDNs play in the cost value chain, and (3) how would the rerouting proposed in this article be actually implemented without terribly impacting performances or previously agreed terms of service. This paper will not answer these entirely, but it provides some elements and hopefully you'll think differently about them after you read it.