An empirical study of the multiscale predictability of network traffic

Abstract

Distributed applications use predictions of network traffic to sustain their performance by adapting their behavior. The timescale of interest is applicationdependent and thus it is natural to ask how predictability depends on the resolution, or degree of smoothing, of the network traffic signal. To help answer this question we empirically study the one-step-ahead predictability, measured by the ratio of mean squared error to signal variance, of network traffic at different resolutions. A one-step-ahead prediction at a coarse resolution is a prediction of the average behavior over a long interval. We apply a wide range of linear and nonlinear time series models to a large number of packet traces, generating different resolution views of the traces through two methods: the simple binning approach used by several extant network measurement tools, and by waveletbased approximations. The wavelet-based approach is a natural way to provide multiscale prediction to applications. We find that predictability seems to be highly situational in practice---it varies widely from trace to trace. Unexpectedly, predictability does not always increase as the signal is smoothed. Half of the time there is a sweet spot at which the ratio is minimized and predictability is clearly the best. Also surprisingly, predictors that can capture non-stationarity and nonlinearity provide benefits only at very coarse resolutions.