Performance optimization in multi-hop wireless mesh networks

Multi-hop wireless mesh networks have emerged as an accepted communication paradigm over the last decade, although their performance is yet to reach to its full potential. Different applications of these networks, such as large-scale metropolitan wireless communication, institutional coverage, battlefield connectivity, etc., often demand highly optimized network performance. To address this issue, this dissertation addresses the problem of performance optimization in multi-hop wireless mesh networks. We maneuver our study on the performance optimization through focusing on different layers in the protocol stack. First, we start at the network layer. We design an energy-efficient network architecture to support high-data-rate applications over low-power wireless meshes, i.e., wireless sensor networks. Here, we study the notion of deploying a high-power radio, 802.11, to augment the network having only conventionally used low-power 802.15.4 radio such that we can guarantee improving the overall energy-efficiency of the network. To achieve the goal, we separately investigate network-level energy efficiencies of both the radios, and then propose the energy-efficient architecture exploiting the findings of the investigation. Our proposed architecture utilizes a delicate balance between the power consumptions of both the radios, which we determine using a cross-layer mathematical model. A real testbed implementation reveals that the proposed network architecture significantly improves the overall energy efficiency of a low-power wireless mesh. Next, we study performance optimization at the transport layer. Here, we specifically focus on reliable data transmission at the transport layer. Our first contribution in this case is based on the observation that the performance of reliable data transmission protocols, such as TCP, depends on the dynamic adjustment of a parameter called retransmission timeout, which in turn depends on the dynamic estimation of round trip time. We propose a stateful Q-learning based technique to dynamically estimate the round trip time. The proposed technique incurs low overhead, making it particularly suitable for networks of embedded systems such as wireless sensor meshes. Besides, the technique substantially improves network performance, which we validate through simulation and real testbed experiments. In addition, we also study on congestion control over wireless mesh networks as the congestion control mechanisms exhibit a substantial influence on overall network performance. We propose end-to-end congestion control mechanisms for both ad-hoc and infrastructure wireless mesh networks. The mechanism for ad-hoc wireless mesh networks is based on the use of a neural network. Besides, the other mechanism for infrastructure wireless mesh networks is based on the use of a decision theory called multi-armed bandit problem. One of the key advantages of these proposed mechanisms is that neither of them involves any radical changes to TCP. Moreover, simulation results and real testbed experiments demonstrate that the proposed congestion control mechanisms can significantly improve the network performance. Finally, we investigate the performance optimization from the upper layers, i.e., the layers that reside on top of the transport layer. Here, we specifically focus on performance optimization over multi-radio networks. First, we study the notion of splitting data, originated from a single flow, over multiple radios. In the study, we propose to perform the splitting from the application layer, as the splitting vastly depends on specific requirements of the application and the application, itself, best knows its own requirements. Besides, we do not have to make any change in existing system architectures in the case of the proposed application-layer data splitting. Moreover, the proposed splitting mechanism exploits a prediction model, and thus incurs no overhead in the network. In addition to such splitting, we also study flow-level data splitting. The notion of flow-level splitting exhibits the potential to improve network performance to a great extent in presence of a mix of different types of flows in the network. In the study on flow-level splitting, first, we analyze coherency between characteristics of the different types of flows and attributes of different radios. Then, we propose a mechanism for flow-level data splitting by exploiting the coherency found from our analysis. We evaluate both of our proposed splitting mechanisms through simulation and real testbed experiments, which uncover their potentials to significantly improve network performance