Improving the performance of distributed simulations of wireless sensor networks.

摘要

Simulations are key to the design, implementation, and evaluation of wireless sensor networks (WSNs) and their applications. To meet the demands for high simulation fidelity and speed, distributed simulation techniques are increasingly being used in WSN simulators. However, existing distributed WSN simulators only provide limited speedup and scalability because of the large overheads in preserving the causality of the interactions of wireless sensor nodes during distributed simulations. In this dissertation, we examine methods to improve the performance of distributed WSN simulators by controlling the overheads related to distributed simulations and parallelizing simulations.;When building distributed simulators, "conservative" and "optimistic" are the two basic approaches for preserving causality. The former ensures causality violations never occur whereas the latter features mechanisms to recover from causality violations. These two approaches incur different overheads and their relative performances vary over different WSNs or simulation hardware. Given that all existing distributed WSN simulators are based on the conservative approach, we study, in the first part of this dissertation, how to improve the performance of the conservative approach in simulating WSNs. We first develop three novel techniques that reduce simulation overheads by exploiting the parallelism in the physical radios, communication protocols and WSN applications. Then we propose a lazy synchronization scheme that further improves simulation performance by identifying and eliminating unnecessary synchronizations during simulations. With these techniques, we implement a fully functional distributed WSN simulator.;In the second part of this dissertation, we study the performance of the optimistic approach in simulating WSNs. Our focus is on understanding the relative performance of the two approaches so appropriate simulation strategies can be devised for a WSN. Since events are handled fundamentally differently across these two classes of simulators, it is difficult to compare the approaches for a specific WSN. We address this challenge by developing a novel trace-based performance evaluation technique that separates simulation overheads from actual simulation algorithms or implementations. This allows one to use the same traces to prototype and evaluate any simulation techniques on virtual platforms with arbitrary hardware. We implement this technique in a simulation performance evaluation framework.