Edge–IoT Computing and Networking Resource Allocation for Decomposable Deep Learning Inference

Abstract

Deep learning (DL) applications have attracted significant attention with the rapidly growing demand for Internet of Things (IoT) systems. However, performing the inference tasks for DL applications on IoT devices is challenging due to the large computational demands of DL models. Recently, edge computing has offered us a solution by deploying resources near the end users. However, resources at the edge are still limited; thus, management issues, such as allocating the networking resources as well as the computing capabilities and configuring the devices appropriately for different applications, become essential. For knobs in such edge management, we consider multiple application tasks with different options of DL models and different hyperparameter settings, along with possible decomposition points that utilize the split DL concept to design the configuration tables. Layer-level decomposition in split DL provides greater flexibility by splitting a single DL inference model into parts on different computing devices, and each part consists of several consecutive layers. We then propose the SplitDL-Image and the SplitDL-Video algorithms based on the Vickrey–Clarke–Groves (VCG) mechanism by considering model performance and frames per second (FPS) requirements with the preferences of the heterogeneous IoT devices. The proposed method allocates networking and edge server computing resources according to the designed configuration tables by assigning the appropriate configuration to each IoT device. Simulation results based on real-world applications show that the proposed method indeed allocates more resources to IoT devices with more urgent/important tasks, preference for better accuracy, or higher local computational cost. In addition, other desired properties, such as truthful bidding, individual rationality, and weakly budget balance, are also guaranteed.