Application Performance Optimization Research Based On Computation Offloading In Mobile Cloud Computing

With the rapid development of mobile Internet technology,mobile devices and mobile applications have penetrated into every aspect of people's lives in recent years.As the development technology and process of mobile applications are gradually matured,and the integration of mobile device sensors is large,it lays the foundation for the birth of a group of applications with great resource requirements.These applications have spawned frequent updates to mobile devices,but the limited resources on them have become a bottleneck limiting applications to provide a better service experience.With the continuous upgrade of mobile networks and the maturity of cloud computing technologies,researchers have attempted to offload the computationally intensive parts of mobile applications to the cloud,which would improve application performance and reduce the power consumption of mobile devices.With the gradual accumulation of results in this research field,people have turned this field into Mobile Cloud Computing(MCC).Although mobile cloud computing has greatly improved the service experience of mobile users,it has also brought new challenges and research problems.After deeply analyzing the relevant research results in the field of mobile cloud computing,this dissertation aims at the performance optimization and carries out the research in the fol-lowing aspects:resource scheduling in multiuser scenarios,transmission scheduling of module-parallel applications,and task scheduling for new CNN mobile applications.First of all,we consider the online resource allocation strategy of cloudlet with limited resources in the multiuser access scenario.We analyze that when multiple users are connected in succession,the online resource allocation scheme presents a se-quence decision process.And the currently resource allocation scheme would affect the resource scheduling of the subsequent requests.Based on the above analysis,we use reinforcement learning to set up online distribution schemes for computing and net-work resources.Specifically,we propose a DQN-based value network to implement the reinforcement learning algorithm.Simulation experiments show that this method has obvious superiority compared with online greedy strategy.Then,this paper analyzes the transmission scheduling problem of the model-parallel application in the distributed execution of the mobile terminal and the cloud after computation offloading.Combined with the computation offloading decision of each module in the application,we designed a two-layer heuristic decision algorithm.The outer layer searches for multiple linear links through depth-first traversal,and determines the execution position of each module on each linear link based on the conclusion of one migration theory.The inner layer first selects the edges that need to be transmitted across the network at the same time,and uses a greedy algorithm to design a data trans-mission strategy.Experiments show that the combination of search pruning and greedy approach would significantly reduce the response time of module-parallel applications under the requirement of satisfying real-time performance.Finally,in order to further improve the inference performance of image recognition applications based on CNN model,this paper proposes a distributed inference framework called IF-CNN based on computation offloading.To reduce the average complexity of the CNN model,we select efficient models from a pool of multiple CNN models with different complexity to process different input images.Specifically,we use multitask learning method to predict the top-1 label probability of the candidate model for the input image,and select the simplest CNN model above a certain threshold based on the value.After selecting the efficient processing model,we implement distributed inference of the CNN model based on the network conditions of the mobile and cloud.In this process,we leverage half-precision reasoning and feature compression to achieve the optimization of two stages of local reasoning and intermediate result transmission.Experiments show that IF-CNN can significantly improve the inference performance of the CNN model without affecting the final recognition rate.At the same time,the model selection process of IF-CNN can complement other inference acceleration methods of CNN models.In this dissertation,aiming at optimizing the performance of mobile applications,we propose a series of optimization schemes to further reduce the response time of applications based on computing offloading.First of all,this paper considers the limited nature of microcloud resources,and uses reinforcement learning to determine online resource allocation schemes for multiuser access.Subsequently,we analyze the char-acteristics of the module parallel application,and design the transmission scheduling strategy after computation offloading.Finally,so as to further improve the inference performance of CNN model,an image-aware distributed inference framework is proposed.We hope that the research work of the paper can provide useful reference and help for application performance optimization in the MCC field.