Communication-aware edge-centric knowledge dissemination in edge computing environments

A colossal amount of data are produced each day and travel through the Internet to be stored or processed on their final destination. Most of them are directly created by people (images, music, videos, etc.), but a considerable amount are generated by sensing and computing devices. Moreover, the Internet of things (IoT) introduces the idea that all devices will have connectivity capabilities, such as connecting to local area networks (LAN) or wide area networks (WAN) including the Internet. By the year 2020, an estimated 30 billion devices will be part of the IoT [1]. This will create the need for sensing and computing devices to communicate efficiently in order to save communication transactions, which will ultimately save communication overhead, resulting in less energy consumption (from not using communication modules like antennas). This chapter proposes an edge-centric predictive methodology, based on real-time model caching, where communication overhead is significantly decreased. This is because only the model’s parameters are cached and disseminated. This event occurs only when the methodology sees a need to update the model. This methodology is more efficient, in terms of communication overhead, compared with continuous raw data transmission. Furthermore, this chapter presents the comparative assessment of the combination of the methodology, with different regression techniques as caching models. Later in this chapter, we will explore the impact of each regression technique on the accuracy as well as the communication overhead. The requirements of implementation are that the end nodes will have at least the processing power and memory of a modern MCU. The achieved results from the analysis suggest that the communication overhead is significantly reduced, by just having a marginally less accurate model. This means that we can have a minor trade-off between accuracy and throughput which can result in improvements in the energy footprint. We provide performance and comparative assessment over real data showing the benefits of the regression models combined with the proposed methodology