Collaborative Inference for Multi-Heterogeneous Edge-End Systems in Intermittent Computing Environments

In intermittent computing environments, multi-edge server collaboration has become a key strategy to optimize task execution in end-edge systems, alleviating the latency and efficiency degradation caused by the resource limitations of a single edge server. However, such collaboration architectures face challenges including server load imbalance and resource allocation inefficiencies arising from heterogeneity and intermittency. To address these issues, this paper proposes a Model Partitioning and Resource Allocation Algorithm in Multi-Server System (MPRA-MSS). The algorithm first identifies the nearest available server based on the distance and resource availability between terminal devices and servers. Leveraging DNN model partitioning techniques, it determines the optimal partitioning point and deploys the deep sub-model on the selected server. Additionally, server resources are allocated based on the data proportion of the deployed deep sub-model, effectively balancing the server load and enhancing system performance. To validate the effectiveness of MPRA-MSS, this study evaluates multiple DNN models using the CIFAR-10 dataset and compares the proposed algorithm with several baseline strategies. Experimental results demonstrate that MPRA-MSS significantly reduces the average task completion time and outperforms existing model partitioning and resource allocation methods.