Data-oriented abstraction of virtual sensors for energy-aware embedded software systems

We analyze the computing and communications as incorporated in networked objects (IoT): such as sensors, with focus on the performance and QoS aspects (e.g., latency of sensor data delivery to end-user). We advocate the offloading of complex computational tasks from the field-deployed low-capability sensor devices to cloud-based remote machines when feasible. In addition to the improved latency performance, the offloading of complex sensor tasks lowers the energy consumption of sensor devices. A key element of our IoT system architecture is the use of layered sensing techniques to determine the offloading of sensor tasks and the network transfer of sensor data. The computational cycles expended and network data transfer overhead vis-a-vis the energy consumption incurred therein, are factored in the partitioning of sensor tasks. Given the large-dimensionality of sensor input data, our architecture incorporates the accuracy and timeliness of sensor outputs as the controllable application-level quality parameters. The paper describes a case study of Optical Character Recognition to corroborate our approach.