Distributed Data Streams

Abstract

We consider a scenario where a server is wirelessly connected to countless sensor nodes that continuously measure data. The task of the server is to monitor the sensors’ data. More precisely, at each time step the server calculates a function defined over the current measurements of the sensors. Since the sensors only have small computational power and tight battery constraints, the communication between the server and the sensors should be as small as possible, i.e., we aim at minimizing the total number of messages that is transferred.There are various conceivable problems for the setting above. We demonstrate our approaches on the following three: In the Top-k-Value Monitoring Problem, the server aims at identifying the k largest values. The Top-k-Position Monitoring Problem shifts the task to identify the sensors observing these values. Finally, the Count Distinct Monitoring Problem obliges the server to determine the number of distinct values currently observed.For all three problems, we not only present communication-efficient protocols for one time step, we also show how it can be exploited if the input at sensors is similar between consecutive time steps to reduce the total communication on the long term. Thereby, we utilize different techniques – involving sampling, dynamic data structures, filter-based approaches, and combinations of them – to demonstrate their potential and their limits in the broad setting described above.