Forecasting models and adaptive quantized bandwidth provisioning for nonstationary network traffic

Network providers are increasingly interested in providing dynamically provisioned bandwidth to customers based on periodically measured nonstationary traffic that meets service level agreements. In this dissertation, we propose a dynamic bandwidth provisioning framework for such a situation. Our framework addresses three critical issues: (1) stochastic property of the traffic may not be available, (2) predicted bandwidth should not over-estimate too much and should avoid under-estimation as much as possible, and (3) frequency of bandwidth updates should factor in overhead incurred due to signaling cost of updates.; For nonstationary periodically measured traffic data, measurements were first collected from Internet access link that connects the University of Missouri-Kansas City to MOREnet. Through statistical analysis, we show that this data set fits a seasonal AutoRegressive Conditional Heteroskedastic based time series model with the innovation process (disturbances) generalized to the class of heavy-tailed distributions.; In generating forecasts from the proposed time series model to accomplish the final goal of bandwidth provisioning, we present two approaches. The motivation for these two approaches arises from the fact the under-forecasts has more impact on the system performance than the over-forecasts. First, we present a modified probability-hop forecasting algorithm that utilizes the confidence-bounds of the conditional forecast distribution to determine the final effective forecast value by expanding a current approach. In the second approach, we introduce a novel generic forecast cost function-based approach that is defined in terms of different penalty functions associated with the under-forecasts and the over-forecasts that also addresses the signaling cost factor. We show existence and uniqueness of the optimal forecast value for this approach under mild assumptions.; For bandwidth estimation, we consider different bandwidth provisioning schemes that allocate or deallocate the bandwidth based on the traffic forecast generated by either of the approaches. These provisioning schemes are developed to allow trade off between the loss and the utilization, while addressing the overhead cost of updating bandwidth. Through extensive studies with three different data sets, we have found that our approach provides a robust dynamic bandwidth provisioning framework for real-world periodically measured nonstationary traffic.