Where in the Internet is congestion

Understanding the distribution of congestion in the Internet is a long-standing problem. Using data from the SamKnows US broadband access network measurement study, commissioned by the FCC, we explore patterns of congestion distribution in DSL and cable Internet service provider (ISP) net- works. Using correlation-based analysis we estimate prevalence of congestion in the periphery versus the core of ISP networks. We show that there are significant differences in congestion levels and its distribution between DSL and cable ISP networks and identify bottleneck sections in each type of network. I. INTRODUCTION Internet congestion has been a topic of active research for as long as the Internet has existed(2). Congestion in computer networks more generally has been studied even longer and is itself predated by research on congestion in telephone networks and car traffic. There are good reasons for being interested in Internet congestion since it largely determines everyday Internet user experience from the time it takes to load a webpage, to the visual quality of streaming media and responsiveness of the on-line gaming experience. One important question concerns the distribution of congestion in the Internet(3). Is it mainly concentrated in the so-called "core" of the network or at its edge or somewhere in between? Is it evenly distributed or concentrated at the edge in some networks and in the core in others? Perhaps congestion can occur in any of those network segments at different times. The answers to these questions have so far proved elusive(2). The companies that own various parts of the network guard their traffic, capacity, and topology data to maintain their competitive advantage, making it virtually impossible for the research community to obtain any kind of insight into the congestion distribution. Yet, even the Internet service providers (ISPs) would benefit from such understanding as it would allow them to target infrastructure improvements at the key points in the network where return on investment, in terms of enhanced user experience, would be greatest. To date, the scarcity of data has been one of the main obstacles to understanding congestion in the Internet. Sim- ulations can shed some light on the range of possible network congestion regimes, but in the absence of real data it is hard to say what is and is not likely to be observed in practice(6). In other words, simulations circumscribe the state space but provide no probability distribution describing the likelihood of various states. This is not surprising because in order to assign such probabilities one would have to know something about the distribution of demand and about routing, which is precisely the information we lack.