Real-time data reduction at the network edge of Internet-of-Things systems

The expected huge increase in the number of IoT data sources (sensors, embedded systems, personal devices etc.) will give rise to network-edge computing, i.e., data pre-processing, local storage, and filtering close to the data sources. Specifically, data reduction at the network edge (e.g., on an IoT gateway device or a mini-server deployed locally at an IoT area network) can prevent I/O bottlenecks, as well as dramatically reduce storage, bandwidth, and energy costs. However, current solutions face two main obstacles towards achieving this benefits of network-edge computing. Firstly, the most efficient algorithms for data reduction of time series (which is one of the prevailing kinds of data in IoT) are developed to work a posteriori upon big datasets and they cannot take decisions per incoming data item. Secondly, the state of the art lacks systems that can apply any of many different possible data reduction methods without adding significant delays or heavyweight re-configurations. This paper presents a solution that automates the switching between different data handling algorithms at the network edge, including an analysis of adjusted data reduction methods, as well as three flavors of a new algorithm that is capable of performing real-time reduction of incoming time series items based on the concept of Perceptually Important Points. The potential benefits are evaluated upon real datasets from street, household, and robot sensors, showing that our solution achieves accuracies between 76,1 % and 93,8 % despite forwarding only 1/3 of the data items, without adding significant forwarding delays.