Abstract
Deep Neural Networks (DNNs) are widely used to analyze the abundance of data collected by massive Internet-of-Thing (IoT) devices. The traditional approaches usually send the data to the cloud and process the DNN inference on the powerful cloud servers but suffer from long network latency. Therefore, edge computing has emerged to reduce network latency by offloading the computation from the cloud to the edge. However, a single resource-constrained edge device is unable to process real-time DNN inference. Thus, we devise a collaborative edge computing system CoopAI to distribute DNN inference over several edge devices with a novel model partition technique to allow the edge devices to prefetch the required data in advance to compute the inference cooperatively in parallel without exchanging data. Subsequently, we present a new optimization problem to minimize the completion time of distributed DNN inference. An innovative algorithm is then proposed to intelligently partition the model into the proper number and sizes of blocks, deploy them on a suitable number of edge devices, and run them in different rounds. The numerical results manifest that our algorithm outperforms the traditional approach by 20%-30% on the completion time.