Research On Key Technologies Of Resource Optimization Management In Edge Networks

With the rapid growth of the network,data at the edge of the network is growing explosively.The use of edge computing technology to lessen the overload load on the cloud and lower the high transmission latency caused by network congestion through the resources and services provided in the edge network has emerged as the primary solution to keep up with the growth of data and ensure the quality of the user experience.However,with constrained resources at the edge,it becomes a challenge to manage edge network resources efficiently in order to increase user quality of experience and lower provider costs.Especially in storage resource management,several challenges arise in caching data and delivery content due to its unique concentration of requested content&dynamic nature,and collaborative caching.To address these challenges,the thesis researches technologies of optimal management for storage resources in Edge Networks,which comprises three research directions.F irstly,accurate prediction of popular content based on request content centralization is proposed,taking into account the transferability of popular content between geographic regions.Secondly,dynamic content updates in real-time are proposed to account for the high-time-varying nature of requested accesses.Lastly,appropriate storage node deployment is proposed based on the collaborative nature of caching and considering transmission delay between storage servers.The main contributions of this thesis are summarized as follows.1.A spatiotemporal-aware popular content prediction algorithm is proposed to address the challenge of predicting short video popular content in scenarios with large data volumes and high time sensitivity.This algorithm maps the influence relationships between regions and regions into graphical form based on the spatial propagation of popular content and creates a graph neural network model.It partitions the input data into three parts:popular content,sub-popular content in the current region,and popular content in other regions.The algorithm periodically trains the weights of the input data of the three parts to improve the accuracy of predicting future popular content of the current region.Furthermore,based on the temporal similarity of popular content,the input data is denoised to reduce a large amount of useless input data and improve training efficiency.Simulation experiments are carried out by using the real access data set of a company.The results show that the accuracy of the proposed prediction algorithm is 3.19%higher than the GCN algorithm,and the training time is reduced by 25.7%compared with the original graph neural network model without denoising processing.2.To address the problem that time-varying access frequency and fixed access frequency threshold affects cache hit rate,a cache replacement algorithm with an adaptive frequency threshold based on user request content concentration information combined with a decision tree algorithm to classify future access frequency is proposed.To address the problem that popular content and non-popular content are difficult to distinguish by access frequency and other features during low-frequency access periods,a cache eviction algorithm based on user concentration information is proposed to capture the relationship between core users and popular content through principal component analysis,and to select eviction content found on this relationship.The experimental results show that the cache replacement strategy can achieve a maximum hit rate of 98%,significantly improving the hit rate of the requested content and thereby providing better user experience quality.3.A cache replacement strategy that integrates admission and eviction is proposed to address the issue of low cache performance caused by the separability of admission and eviction in cache replacement policies of edge networks.Firstly,to improve the cache hit rate,two modules,namely delayed eviction and unified standards,are added to the existing cache replacement algorithms to address the issues caused by the separability of admission and eviction,such as the gap between the retrieved information and the content value.Then,a new content value function is defined by fully utilizing the request information generated during retrieval.This function combines the two modules to form an admission and eviction integrated cache replacement strategy based on the A3C algorithm.Simulation experiments are conducted using a specific company’s user access log information.According to experimental findings,the hit rate of the proposed strategy is 2.7 times greater than LRB and reduces transmission delay by a maximum of 37.5%.4.A storage node deployment strategy based on access patterns is proposed to address the transmission delay problem among storage servers caused by the collaborative caching of edge servers.First,a new criterion based on the diffusion capacity of edge nodes is defined to measure the server transmission delay,which is more accurate than existing metrics.Then,based on this criterion,the problem is modeled as a minimization of the number of storage nodes.An improved genetic algorithm is proposed to account for the non-uniform server request frequency and the non-linear relationship between the number of servers and transmission delay.Simulation results show that this deployment strategy can reduce the number of storage nodes and deployment costs while maintaining the same user tolerance for delay.