In recent years, with the development of deep neural network (DNN), more and more applications (e.g., image classification, target recognition and audio processing) are supported by it. However, the disadvantage of its own large model makes it difficult to apply on resource-constrained devices such as mobile devices. In order to solve this problem, the existing research and technology mainly focus on the DNN model compression and the segmentation migration of the model. The former is generally at the expense of reducing accuracy, and the segmentation of the model has no unified migration tool for the DNN model of different applications. In this work, we propose a universal neural network layer segmentation tool, which enables the trained DNN model to be migrated, and migrates the segmentation layer to the nodes in the current network in accordance with the dynamic optimal allocation algorithm proposed in this paper. The experimental results show that the tool can adapt to various neural networks with different structures and perform optimal allocation of layers through algorithm. When the number of working nodes increases from 1 to 5, this method can speed up DNN 2-2.5 times, and shows a good acceleration effect.

The internet of things (IoT) attracts great interest in many application domains concerned with monitoring and control of physical phenomena. IoT applications try to provide more and more functionality and then they inevitably become so complex as to make the limits of devices worse, which may lead to poor performance of applications. Computation offloading is a promising way to improve the performance of an IoT application by executing some parts of the application on remote devices or servers. However, supporting such capability is not easy for application developers due to (1) adaptability: IoT applications often face changes of runtime environments so that the adaptation on offloading is needed. (2) effectiveness: when the device context changes, it needs to dynamically decide the deployment plan of computation tasks, and the reduced execution time must be greater than the network delay and extra overheads caused by offloading. This paper proposes a framework which supports IoT applications with adaptive computation offloading capability. First, a design pattern is proposed to enable an application to be computation offloaded on-demand. Second, an estimation model is presented to automatically decide the deployment plan for offloading. Third, a framework is implemented to support the design pattern and the estimation model. A thorough evaluation on the real-world application is proposed, and the results show that our approach can help reduce execution time by over 45% in most scenarios.