End-to-End Internet Packet Dynamics

This very detailed paper uses traces of TCP connections between monitored hosts on the Internet to characterize latency, throughput, loss rates, reordering, corruption rates, etc for Internet paths. The most interesting part is the methodology - gleaning information from the highly noisy traces through careful reasoning about the queueing dynamics involved. For example, I liked how PBM improved over packet pair by allowing measurement of multiple bottlenecks and estimating how many channels are sharing a bottleneck. Some of the lessons in the paper, such as prefering measurements at the receiver, are useful for protocol designers (who must estimate RTT, bandwidth, etc) and other researchers interested in Internet measurement.

The main conclusion of the paper is that Internet conditions are highly variable from path to path and even over time on a single path, and that weird things like reordering, duplication, corruption, etc do happen. Paths are also highly asymmetrical. These are all very good observations to put forward so that the research community is not guided by overly simplistic models when evaluating new Internet protocols. The only unfortunate thing is that although this paper tells us the variability of conditions involved, it does not tell us how we might build a good model for evaluating protocols.

My main question after reading this is, has work like this been done for today's Internet too? It would be especially interesting to see how the dynamics have changed with the introduction of broadband, the expansion of the Internet into Asia and Africa, and the introduction of mobile hosts (e.g. some of the hosts could be chosen to be devices rather than servers at corporations and universities).