Geelytics: Enabling On-Demand Edge Analytics over Scoped Data Sources

Abstract

Large-scale Internet of Things (IoT) systems typically consist of a large number of sensors and actuators distributed geographically in a physical environment. To react fast on real time situations, it is often required to bridge sensors and actuators via real-time stream processing close to IoT devices. Existing stream processing platforms like Apache Storm and S4 are designed for intensive stream processing in a cluster or in the Cloud, but they are unsuitable for large scale IoT systems in which processing tasks are expected to be triggered by actuators on-demand and then be allocated and performed in a Cloud-Edge environment. To fill this gap, we designed and implemented a new system called Geelytics, which can enable on-demand edge analytics over scoped data sources via IoT-friendly interfaces to sensors and actuators. This paper presents its design, implementation, interfaces, and core algorithms. Three example applications have been built to showcase the potential of Geelytics in enabling advanced IoT edge analytics. Our preliminary evaluation results demonstrate that we can reduce the bandwidth cost by 99% in a face detection example, achieve less than 10 milliseconds reacting latency and about 1.5 seconds startup latency in an outlier detection example, and also save 65% duplicated computation cost via sharing intermediate results in a data aggregation example.