CNN-DAG-Editor: A Convolutional Neural Network offloading analyzer with Multi-Objective Dynamic Adaptive Resource Competitive Swarm Optimization

With the rapid development of artificial intelligence applications on mobile devices, there are increasing demands for optimizing the runtime, energy consumption, and cost-effectiveness of Convolutional Neural Networks (CNNs). These objectives often cannot be simultaneously optimized in real-world applications. The most effective way to enhance CNN performance on mobile devices is through CNN offloading while existing research often considers only a single network architecture with a single optimization objective, without addressing runtime, energy consumption, and cost-effectiveness as a multi-objective optimization problem. In this paper, we propose a CNN offloading analysis tool called CNN-DAG-Editor and introduce a Multi-Objective Dynamic Adaptive Resource Competitive Swarm Optimization (MDARCSO) algorithm within CNN-DAG-Editor for optimizing CNN offloading across devices, edge servers, and cloud servers. Experiments show that our Edge-Cloud-Server Collaborative Offloading (ECESOPS) strategy, based on MDARCSO, outperforms other strategies like No Offloading Policy (NOPS), Cloud-Server Full Offloading Policy (CFOPS), and Hybrid Offloading Policy (HOPSO) in terms of fitness performance, task energy consumption, and leasing costs. Furthermore, to verify the performance of the MDARCSO algorithm, we compared it with six state-of-the-art LSMOEA algorithms on a public benchmark (LSMOP). The results demonstrate that MDARCSO achieves the best overall performance on LSMOP.