Joint Optimization of the Partition and Scheduling of DNN Tasks in Computing and Network Convergence

Abstract

Computing and network convergence (CNC) is a new network architecture based on computing evolution and network integration. Deep Neural Networks (DNNs) inference imposes a heavy computational burden on mobile devices. In this letter, an end-edge-network-cloud (EENC) collaborative inference architecture is proposed to reduce the DNN inference latency and maximize the computing potential of the CNC. A heuristic Centralized DNN Task Offloading algorithm (CDTO) is proposed for the fine-grained partition and scheduling problems of multiple DNN inference tasks. The CDTO algorithm can significantly reduce the makespan of DNN inference tasks and effectively improve the concurrent capacity of DNN tasks.