A Model-based Framework for Autonomic Performance Management of Wireless Mesh Networks.

摘要

While performance management in wireless mesh networks is already non-trivial, the addition of quality-of-serviced based traffic creates an even more challenging task. We propose an Autonomic Network Performance Management (ANPM) Framework, which includes four functional components: monitoring, adaptive modeling, optimization, and configuration units. These components interact to determine the optimal set of controllable factor values that will (i) maximize system-wide performance or (ii) meet a specified performance target. The proposed ANPM architecture is a feedback-based adaptive controller that can be implemented on each network node in a fully-distributed approach, on a centralized server, or as a hybrid approach based on a clustering of network nodes.;The monitoring unit is responsible for monitoring/gathering relevant factors and metrics information that defines the current network and channel state. This information is then passed to the adaptive modeling and performance optimization units. The selection of factors and metrics is determined by the adaptive modeling unit and is critical to the overall performance of the ANPM system and governs the robustness of network model.;The goal of the Adaptive Modeling Unit is to develop empirical-based models that effectively characterize the network dynamics. We compare between different types of models (e.g., linear regression and non-linear neural networks) with respect to their accuracy and complexity. We use statistical Design-of-Experiment and Analysis-of-Variance to gather and screen all data to determine significant factors. We propose a ranking table methodology that determines the best set of factor values for a given network state in O(n logn) where n is the number of network factors considered. We also propose a technique to automate the learning and extraction of network rules that are used by ANPM to better optimize network state. Given the results of the optimization processes, the configuration unit initiates the necessary protocol and parameter reconfiguration. We present two simulation-based case studies that illustrate network reconfiguration by changing entire network protocols and adjusting protocol parameter values. Finally, we present hardware implementation to explore the feasibility of real-time protocol switching and provide insight to key issues, such as switching time, effect of traffic load on switching time, and performance improvement after reconfiguration.