Joint DNN Partition and Resource Allocation Optimization for Energy-Constrained Hierarchical Edge-Cloud Systems

Hierarchical edge-cloud systems collaboratively utilize the resources of both the edge server and central cloud, enabling deep neural network (DNN) partition between the edge and cloud to accelerate the inference. However, the limited energy budgets of both the edge server and central cloud restrict them from providing optimal DNN inference services. Moreover, considering the high dynamics in stochastic environments, the long-term system performance should be optimized under long-term energy constraints. How to improve the long-term DNN inference performance in such energy-constrained hierarchical edge-cloud systems is less studied by existing related works. In this paper, we aim to jointly optimize DNN partition and computing resource allocation to minimize the long-term average end-to-end delay of multiple types of deep learning (DL) tasks while guaranteeing the energy consumption of the edge server and central cloud within their energy budgets. Based on the Lyapunov optimization technique and reinforcement learning, we design a novel deep deterministic policy gradient based DNN partition and resource allocation (DDPRA) algorithm to train policy to decide DNN partition dynamically by observing the environment. Moreover, the DDPRA algorithm is embedded with a heuristic computing resource allocation (HCRA) algorithm, which effectively reduces the complexity of policy training by decoupling and optimizing the computing resource allocation separately. We analyze the complexity of our algorithms and conduct extensive simulations. The numerical results demonstrate the superiority of our algorithm in comparison with 5 other schemes in multiple scenarios.