A Structural Analysis of Network Delay

Network delay is a crucial metric for evaluating the state of the network. We present in this paper a structural analysis of network delay, based on delay measurements of aback bone network. This delay analysis is performed using a subspace method called Principal Component Analysis (PCA). The analysis reveals that the delay time series can be decomposed into two constituents: a smooth periodic trend and a set of sparse bursts. We call the former the "normal" component and the latter the "abnormal" component. While this structural decomposition is appealing and may be used to further infer other delay information of interest, we find that using PCA in delay analysis has the same challenges as used in traffic analysis. Particularly, it experiences performance degradation due to the so called "perturbation phenomenon".