Edge-Distributed IoT Services Assist the Economic Sustainability of LEO Satellite Constellation Construction

Abstract

There are thousands or even tens of thousands of satellites in Low Earth Orbit (LEO). How to ensure the economic sustainability of LEO satellite constellation construction is an important issue currently. In this article, we envision integrating the popular and promising Internet of Things (IoT) technology with LEO satellite constellations to indirectly provide economic support for LEO satellite construction through paid IoT services. Of course, this can also bring benefits to the development of IoT. LEO Satellites can provide networks for IoT products in areas with difficult conditions, such as deserts, oceans, etc., and Satellite Edge Computing (SEC) can help to reduce the service latency of IoT. Many IoT products rely on Convolutional Neural Networks (CNNs) to provide services, and it is difficult to perform CNN inference on an edge server solely. Therefore, in this article, we use edge-distributed inference to enable the IoT services in the SEC scenario. How to perform edge-distributed inference to shorten inference time is a challenge. To shorten the inference latency of CNN, we propose a framework based on a joint partition, named EDIJP. We use a joint partition method combining data partition and model partition for distributed partition. We model the data partition as a Linear Programming (LP) problem. To address the challenge of trading off computation latency and communication latency, we designed an iterative algorithm to obtain the final partitioning result. By maintaining the original structure and parameters, our framework ensures that the inference accuracy will not be affected. We simulated the SEC environment, based on two popular CNN models, VGG16 and AlexNet, the performance of our method is varified. Compared with local inference, EdgeFlow, and CoEdge, the inference latency by using EDIJP is shorter.