Parallel Detection of Temporal Events from Streaming Data

Abstract

Advanced applications of sensors, network traffic, and financial markets have produced massive, continuous, and time-ordered data streams, calling for high-performance stream querying and event detection techniques. Beyond the widely adopted sequence operator in current data stream management systems, as well as inspired by the great work developed in temporal logic and active database fields, this paper presents a rich set of temporal operators on events, with an emphasis on the temporal properties and relative temporal relationships of events. We outline three temporal operators on unary events (Within, Last, and Periodic), and four ones on binary events (Concur, Sequence, Overlap and During). We employ two stream partitioning strategies, i.e., time-driven and task-driven, for parallel processing of the temporal operators. Our analysis and experimental results with both synthetic and real-data show that the better partitioning scheme in terms of system throughput is the one which can produce balanced data workload and less data duplication among the processing nodes.