Design And Evaluation Of Incremental Learning Based Edge Caching Algorithms

In recent years,with the rapid development of various new mobile Internet applications and data services,edge network data caching has become an important way to cope with the rapid growth of network data traffic.By transferring the data traffic to the edge network and improving the performance of content cache replacement algorithm,the data storage is closer to the user and the network resource access performances are effectively improved.Besides,the backhaul link bandwidth pressure is reduced and the user experience of mobile Internet is improved.By making full use of the computing,storage and communication capabilities of edge network nodes,this paper studies the edge cache management algorithm based on machine learning.By learning the request pattern,it predicts the hot contents and makes caching replacement decisions,so as to improve the cache hit rate of edge nodes.Due to the mobility of user access,the tidal nature of different types of content requests,and the global load balancing,the user request pattern will continue to change.At present,caching replacement algorithms based on machine learning have two technical challenges: on the one hand,general machine learning algorithms need to retrain the model regularly to cope with the changes of user request patterns,which is difficult to effectively cope with the changes of user request patterns;on the other hand,due to the different cache capacities of edge nodes,different caching admission and eviction strategies have different impact on the hit rate of edge nodes.The main research content of this paper includes the following two aspects.1)An incremental learning based caching admission algorithm is designed.Firstly,the prototype of edge caching system is constructed,and the method of extracting object popularity features from request flow is designed,and the impact of different training features on cache hit rate is compared;the method of dynamically generating training samples from request flow to update incremental learning model is designed,which retains previously learned knowledge and learns new knowledge from limited incremental data to adapt to user request.Next,the method predict hot objects and admit into the cache according to the incremental learning algorithm.Through the experiments on real world data sets,the incremental learning based caching admission algorithm improves the hit rate of edge cache system.2)An incremental learning based caching eviction algorithm is designed.With the increase of cache capacity,the improvement of cache hit rate of the incremental learning based caching admission algorithm tends to be flat.To improve this problem,an incremental learning based caching eviction algorithm is designed for large cache capacity scenarios.This section designs a method to extract the access characteristics of objects in cache and establishes standard of cache eviction.Then,this section designs a method to sample objects from the cache,construct training samples,and update the incremental learning model.By predicting whether the cache objects will be accessed within a time limit,the method retains or evicts some objects from the cache.Through the verification on real world data sets,the incremental learning based caching eviction algorithm improves the cache hit rate in large cache capacity scenarios.