UC San Diego

In this thesis, we identify patterns of node-lifetimes among nodes participating in PlanetLab. While it is common to assume that time-to-failure follows an exponential distribution, we show that this assumption does not hold true in trace data collected from the PlanetLab test bed. By applying clustering techniques to over a year's worth of time-to-failure data from PlanetLab, we identify a set of six node lifetime distributions that are well-suited to characterizing the nodes of this particular dataset. Because of the uniqueness of the distributions, the characterization provides the additional benefit of allowing us to infer anomalies such as common administrative policies, outages, and other such events