Design and self-management of wireless networked systems with model-driven optimization

摘要

Wireless networked systems are becoming increasingly popular, with a growing number of deployments and diverse applications. This leads to increasing complexity of these systems, as due to the shared nature of the wireless medium the amount of network resources available is constantly decreasing. These networks have to operate in challenging and dynamic conditions, and accommodate diverse and often contradictory objectives stated by multiple stakeholders. Often wireless networks are integral parts of larger systems, like the Internet, and need to utilize existing imperfect software and hardware components, which results in additional heterogeneity and implementation challenges. All this makes design, planning, and run-time management of wireless networked systems a demanding task. In this thesis we address some of these challenges through a model-driven optimization methodology, which is formalized using category theory. We state a meta-optimization problem based on requirements of network players, operational context, applicable models that consider for both parameter- and component-based solutions. In order to simplify the problem we propose applying the mapping functions of abstraction, transformation, and decomposition, while considering the aspects of information loss or disruption in the problem formulation. We show on selected case studies how the proposed methodology can be employed at both static (design, planning) and dynamic (run-time management) stages of a network's life-cycle. In particular, we demonstrate that this methodology is effective for a detailed design and implementation of a self-optimizing system for wireless home networks, and optimization of protocol stacks for wireless sensor systems with an ontology-driven framework. Since modeling is a crucial aspect of network optimization, we also propose and investigate several types of models in this work. Directed labeled graphs and their network motifs are used for identification of the networking context in the CSMA/CA based networks. We show how spatial network structure affects characteristics of these graphs, such as node degree distributions and occurrence patterns of small network motifs. We also propose a novel graph edge labeling based on clustered correlation coefficients that capture network dynamics imposed by tunable network parameters. We argue that this robust metric can lead to faster network optimization in a variety of operational conditions. We also consider modeling of the temporal context. In particular, based on power spectrum measurements we obtain online hidden semi-Markov models of network activity patterns and apply them as part of a dynamic spectrum access scheme. Additionally, we exploit metaheuristic mechanisms for cross-layer optimization and network planning that aim at robustness, performance maximization and optimizability of the system. We investigate how the size and the structure of the state space, the availability of valid models and utility functions influence the convergence of these methods. Our results show that for network problems it is often better to invest in careful problem formulation and long execution of simple metaheuristics rather than going for their custom modifications besides the simplest ones. The proposed approaches have being extensively prototyped or proven through simulation based experimentation. In particular we have focused on small-scale wireless networked systems that utilize IEEE 802.11 radio interfaces and wireless sensor networks. We have also experimented with the WARP software defined radio platforms.