Hierarchical and predictive design techniques for improving QoS and performance in modern wired and wireless networks.

摘要

In this dissertation, we first propose and evaluate a hierarchical totally mobile wireless architecture (HTMWA) that combines the advantages of the ad-hoc and the cellular models. Our tests have shown that when the total number of channels is kept the same, the two-tier system outperformed the one-tier counterpart under all load conditions. Under the constraint of equal power consumption, the two-tier system still achieved improvement over the one-tier system, especially at light and medium load levels. We introduced load balancing schemes based on the concept of reversible handoffs and analyzed the resulting performance improvement. We also investigated different channel splitting schemes that aim at optimizing the distribution of the limited spectrum resources among the mobile routers in the two tiers of the hierarchy. An analytical model to compute the new call and handoff blocking probabilities in a two-tier HTMWA has been developed and validated. To enhance the survivability of HTMWA, we developed and evaluated several recovery protocols to deal with the loss of mobile routers.; We conclude the dissertation by proposing and evaluating a new linear traffic predictor that can be used for dynamically resizing the bandwidth of VPN (virtual private network) links. We present the results of extensive performance comparisons of three known predictors: Gaussian, ARMA (auto-regressive moving average) and fARIMA (fractional auto-regressive integrated moving average). Our tests have shown that the simple Gaussian predictor with higher mean square error (MSE) often outperforms the ARMA and fARIMA predictors with lower MSEs. Guided by our performance tests, we developed a new predictor for link resizing: L-PREDEC (linear predictor with dynamic error compensation). Our performance tests have shown that L-PREDEC works better than Gaussian, ARAM and fARIMA in terms of the three metrics listed above. Finally, we investigated how to choose the best L-PREDEC refitting period for non-stationary long traffic traces. We showed that the optimal refitting period should be long enough to smooth out local burstiness and short enough to filter out the long-term changes (trend and seasonality) in the traffic rate.