On robust and efficient performance over hybrid wireless networks

In the previous decade, we have witnessed an explosion in wireless internet access. Currently, two dominant solutions are cellular data services and WiFi (WLAN Hotspot) services. The Hybrid Wireless Network model is a recent proposal in which wireless devices are dual interfaced, capable of detecting and switching between the cellular and WiFi network services, so that the advantages of the both services can be explored. In this dissertation, we identify several performance issues in Hybrid Wireless Network model from an end user's perspective, and study how to achieve robust and efficient network operations to improve the overall service quality.;We start from the performance of packet scheduling in a cellular network system. The state of the art scheduling algorithms typically adopt an adaptive approach which depends on the channel prediction in next time slot. When the channel error exhibits a wide range of patterns and the prediction is imperfect, such algorithms perform poorly. In this dissertation, we derive a statistical lower bound for scheduling under uncertain channel conditions. In particular, given a coarse-grain channel prediction, we study what is the robust scheduling strategy for the base station to achieve worst case optimal performance from a zero-sum game formulation.;Recently, multi-hop relaying has been proposed to improve the overall cellular service quality. In this dissertation, we further study the impacts of such multi-hop operation to the end to end transport performance. We have made two contributions in making TCP more efficient over the multi-hop wireless networks. First, we show that given a multi-hop path, there exists a best sending window with which TCP achieves its highest throughput. Using its current congestion control mechanism however, TCP tends to overload the network resulting in sub-optimal performance. We formally establish a model between the contention packet loss and network load. Two link-layer mechanisms are introduced to help TCP stabilize its transmission window around the optimal operating point that allows for maximum spatial reuse along the delivery path.;Second, multi-hop delivery path is more susceptible to node mobility and channel errors than single hop WLAN. In order to achieve good end-to-end throughput, TCP has to estimate the current network conditions such as transient network disconnection, channel errors or congestions, and respond differently. We show that it is possible to construct a set of TCP level metrics with which network conditions can be detected without the assistance of the intermediate nodes. The idea is to apply cross validation among multiple end-to-end metrics to filter out measurement noises and reduce the false alarms. We implemented the proposed ADTCP in both NS-2 simulator and Linux ipv4 kernel. From the simulations and experiments, we show that in scenarios where channel errors and node mobility are introduced, ADTCP outperforms TCP NewReno significantly.