The widespread application of IoT (Internet of Things) systems in industries brings complex and high-precision computing requirements. Applying Machine Learning programs, typically Neural Networks, on-device can meet the massive computation needs but challenges compact energy sources of IoT devices. To save the device's energy consumption, this paper builds an edge-cloud distributed CNN implementation with serverless computing, which requires no configuration, update, and scaling like cloud-server-based solutions. To characterize edge-cloud CNN inference under the serverless computing environment, we conduct systematic evaluations of our solution on an IoT platform on energy consumption, resource cost, and execution latency. By testing a convolutional neural network and a residual network, this serverless schema can reduce the device's energy consumption by up to 41.8% and 63.3% for AlexNet and ResNet, respectively, compared to the entire edge CNN inference. Even it achieves an 18.6% energy reduction at most for AlexNet when comparing with the complete cloud execution. As for executing performance, we have also analyzed the time overhead under the fine-grained time fraction. The results show that long cloud response latency and data access delay on-cloud are the main performance bottlenecks in the serverless-computing-based edge-cloud CNN execution.
Wang et al. (Thu,) studied this question.