Disjoint Parallelization of Sliding-Window Streaming Aggregation

Abstract

Online analysis of data streams (with different degrees of parallelism) is becoming progressively more important as the amount of sensory data is growing. Aggregate functions represent one common example thereof. The current parallel approaches for online aggregation of data streams share one characteristic: one or several threads serve as central coordinators by distributing the work as well as the incoming data to other threads dedicated to its processing. If an approach uses centralized coordination units, its scalability is bounded to the throughput with which such coordination units can distribute work and data to processing threads. This thesis deals with the development of online analysis approaches for streaming aggregation without centralized coordination units. The coordination tasks as well as the remaining work are distributed among all participating threads. In this thesis we first introduce basics of online analysis and streaming aggregation, and then we provide several options for disjoint parallelizations of sliding-window based streaming aggregation. We study the developed approaches' runtime properties in order to maximize their throughput and minimize their latency. We also evaluate their throughput and latency in practice and discuss related work that could improve certain aspects of the approaches developed within this thesis.