A mixture model for the connection holding times in the Video-on-Demand service

A fundamental traffic parameter for the design of a network supporting the Video-on-Demand service is the connection holding time. This holding time is usually characterized by its average value and, possibly, by its probability density function (pdf). In this paper, a model is proposed for this density, consisting of a mixture of Gaussian pdf’s, whose parameters are estimated through the combination of the K-means algorithm and of a recursive likelihood maximization procedure. The model has been applied to a data set extracted from a video catalogue. Exponential and log-normal models have been fitted to the same data set by comparison. For all three models, the chi-square statistic and the Akaike Information Criterion result to be much lower for the mixture model than for the other two. In addition, the shape of the mixture pdf resembles more closely the multimodal characteristic of the empirical data. This results have been obtained with a mixture of limited complexity, being made up of a small number of Gaussian components (three or four).