Model Placement Algorithm for Distributed DNN Inference with Collaborative Edge Computing

With the increasing complexity of deep neural networks (DNN) models, traditional single-device solutions struggle to meet the requirements of real-time inference tasks. Edge computing, enabling edge-edge collaboration, has emerged as a new paradigm for distributed DNN inference. However, due to the complex inter-layer dependencies in DNN models, the quality of the model placement strategy across edge servers significantly affects the inference latency. To this end, this paper proposes an inference placement algorithm for DNN models to minimize the total latency. First, we construct a distributed inference system model involving multiple edge servers and formulate an overall inference latency minimization problem. Second, we design a branch-based dynamic programming algorithm that updates edge server selection strategies for individual branches based on the input-output mapping of DNN layers. Furthermore, we propose a greedy global integration algorithm that prunes candidate solutions across multiple branches to reduce computational complexity. Simulation experiments demonstrate that, compared to other algorithms, our algorithm reduces inference latency by 7.2% and 11.6% on AlexNet and GoogleNet, respectively, while exhibiting stronger adaptability under varying bandwidth and computing power conditions.