Dynamic Bandwidth Allocation for Video Traffic Using FARIMA-Based Forecasting Models

In this work time series forecasting models and techniques are implemented to video traffic as part of three dynamic bandwidth allocation schemes. Traffic produced by videos is known to exhibit characteristics such as long and short range dependencies but as it is shown here non-linearity and conditional volatility may also appear as potential characteristics and then affect the choice of forecasting techniques. While models such as FARIMA, ARIMA and Holt-Winters have been used as traffic predictors in bandwidth allocation schemes, we attempt to improve the accuracy of video traffic predictions by using FARIMA/GARCH, hybrid FARIMA or FARIMA/GARCH with neural networks, a model selection strategy based on a non-linearity test, and a forecasting strategy which combines the forecasts produced by a FARIMA, a FARIMA/GARCH and a neural network. The traffic forecasts are used to allocate bandwidth following three different dynamic schemes. The performance of the different forecasting approaches is then tested on eight traces, aggregated on different timescales (frames, GoPs or seconds); and their comparison pertains their predictive capacity but mainly their cost effectiveness when contributing to dynamic bandwidth allocation approaches. Lastly, using the best forecasting approach, which on the average appears to be the hybrid FARIMA/GARCH-MLP model it is possible to evaluate the allocation schemes based on buffering and utilization rate, average and maximum queue length and total number of changes of allocated bandwidth.