Measurement and Modelling of the Temporal Dependence in Packet Loss

Understanding and modelling packet loss in the Internet is especially relevant for the design and analysis of delay-sensitive multimedia applications. In this paper, we present analysis of hours of endto-end unicast and multicast packet loss measurement. From these we selected hours of stationary traces for further analysis. We consider the dependence as seen in the autocorrelation function of the original loss data as well as the dependence between good run lengths and loss run lengths. The correlation timescale is found to be or less. We evaluate the accuracy of three models of increasing complexity: the Bernoulli model, the 2-state Markov chain model and the -th order Markov chain model. Out of the trace segments considered, the Bernoulli model was found to be accurate for segments, the 2-state model was found to be accurate for segments. A Markov chain model of order or greater was found to be necessary to accurately model the rest of the segments. For the case of adaptive applications which track loss, we address two issues of on-line loss estimation: the required memory size and whether to use exponential smoothing or a sliding window average to estimate average loss rate. We find that a large memory size is necessary and that the sliding window average provides a more accurate estimate for the same effective memory size.