Learning-aided fine grained offloading for real-time applications in edge-cloud computing

Edge-cloud computing has been widely adopted to provision abundant resources for latency-sensitive and computation-intensive vehicular applications in Internet of vehicles (IoV), bringing more entertainment, security, and efficiency on the road. Generally, the applications in real world are composed of massive subtasks with dependent relationships which are commonly modelled as directed acyclic graphs (DAGs), and thus fine grained offloading and parallel computing are imperative during offloading to promote the quality of service. However, due to the diversity of DAG-based applications and the complexity of dynamic edge-cloud environment, the vehicle intelligent management system is incapable of scheduling offloading with effect, resulting in additional transmission latency and energy expenditure on wireless channels and backhaul links. To reduce application response time and meanwhile save the energy consumption, a markov decision process is formulated based on the fine grained offloading with the intention of obtaining an optimal policy. Besides, to make offloading more adaptive to various application scales, a learning-aided fine grained offloading for real-time applications, named LFGO, is designed with deep q-learning in edge-cloud empowered IoV. Eventually, experiments are conducted with generated DAGs based on real-world applications, covering a wide range of subtask numbers, transmission rate and computing capability, to verify the efficiency of LFGO.