Calibration and Analysis of Enterprise and Edge Network Measurements

Abstract

With the growth of the Internet over the past several decades, the field of Internet and network measurements has attracted the attention of many researchers. Doing the measurements has allowed a better understanding of the inner workings of both the global Internet and its specific parts. But undertaking a measurement study in a sound fashion is no easy task. Given the complexity of modern networks, one has to take great care in anticipating, detecting and eliminating all the measurement errors and biases.

In this thesis we pave the way for a more systematic calibration of network traces. Such calibration ensures the soundness and robustness of the analysis results by revealing and fixing flaws in the data. We collect our measurement data in two environments: in a medium-sized enterprise and at the Internet edge. For the former we perform two rounds of data collection from the enterprise switches. We use the differences in the way we recorded the network traces during the first and second rounds to develop and assess the methodology for five calibration aspects: measurement gain, measurement loss, measurement reordering, timing, and topology. For the dataset gathered at the Internet edge, we perform calibration in the form of extensive checks of data consistency and sanity.

After calibrating the data, we engage in the analysis of its various aspects. For the enterprise dataset we look at TCP dynamics in the enterprise environment. Here we first make a high- level overview of TCP connection characteristics such as termination status, size, duration, rate, etc. Then we assess the parameters important for TCP performance, such as retransmissions, out-of-order deliveries and channel utilization. Finally, using the Internet edge dataset, we gauge the performance characteristics of the edge connectivity.