Flow Vector Prediction in Large IP Networks

This paper considers the problem of predicting the number, length and distribution of IP traffic flows some time into the future, based upon packets collected in the present. A particle filter is used to predict the mean flow length and complete flow distributions for subsequent timesteps. A model for the histogram of flows corresponding to any given time interval is first presented, and the particle filter is then used to estimate the parameters of the model. The proposed method was tested on a large number of commonly-available data traces, and the results were analyzed in terms of the difference between the predicted flow distributions and actual flow histograms. An important application of this work is in resource reservation for protocols that require guaranteed qualities of service.