Abstract: Recent years have witnessed a new class of monitoring applications that need to continuously collect information from remote data sources. Those data sources, such as web click-streams, stock quotes, and sensor data, are often characterized as fast-rate high-volume ``streams''. Distributed stream-processing systems are thus designed to efficiently use system resources to serve the data-acquisition needs of the applications. Most of the state-of-the-art stream-processing systems assume an Ethernet-based network whose bandwidth is abundant, and focus on mechanisms to save computational power and memory. For applications involving wireless networks, particularly multi-hop mesh networks, we recognize that the most limiting factor in efficiently processing streams lies in the network's highly constrained bandwidth. Hence, this dissertation proposes a group-aware stream filtering approach that saves bandwidth at the cost of increased CPU time, for low-bandwidth data-streaming systems. This approach, used together with multicasting, exploits two overlooked properties of monitoring applications: 1) many of them can tolerate some degree of ``slack'' in their data quality requirements, and 2) there may exist multiple subsets of the source data satisfying the quality needs of an application. We can thus choose the ``best alternative'' subset for each application to maximize the data overlap within the group to best benefit from multicasting. After proving the problem NP-hard, we introduce a suite of heuristics-based algorithms that ensure data quality, specifically data granularity and timeliness, in addition to preserving network bandwidth. Our framework for group-aware stream filtering is extensible and supports a diverse range of filtering needs of monitoring applications. We evaluate this approach with a prototype system based on real-world data sets. The results show that quality-managed group-aware filtering is effective in trading CPU time for bandwidth savings, compared with self-interested stream filtering. We also evaluate the effect of each algorithm on temporal freshness of the data. Finally, we discuss other application realms that might benefit from group-aware stream filtering.
Copyright © 2008 by Ming Li.