Research On Heuristic Proactive Caching Strategy For Edge Computing

Recent years have witnessed tremendous growth of wireless mobile devices,which has yielded unprecedented traffic demands.These transmission demands have placed a huge load on networks,resulting in network congestion and high transmission delays,which have reduced the quality of user experience.To satisfy the considerable amount of mobile application requests,one promising approach is via the deployment of small cell networks(SCNs).That is,to deploy a short-distance and low-power secondary base station in the lower layer of the macro-cellular network.User devices in radio range can connect to small base stations,thereby increasing spatial reuse and network capacity.In actual,the rapid rise of mobile content access has led to a dense deployment of wireless networks especially in urban environments.Since the large-scale deployment of SCNs hinges on physical space and expensive cost,the improvement of this method is limited.Edge computing leverages proximate infrastructure to compute or store.Content is expected to be stored in small base stations to shorten transmission distances and alleviate backhaul congestion.The distribution of content popularity is generally modeled as a Zipf distribution,that is,a fraction of content accounts for the vast majority of the traffic load.Therefore,we propose heuristic adaptive proactive caching strategy(APCS)based on the prediction of content popularity for edge computing.The paper first introduces the temporal and spatial correlation of content popularity,then proposes the architecture of mobile edge computing(MEC),which includes mobile user devices,edge nodes,and the cloud server.User devices are the generator of the request tasks,which is not only the consumer of data but also the producer.Edge nodes are service clusters close to users,which provide computing and storage capabilities,including macro base stations and small base stations.The cloud server is similar to the central service cluster in cloud computing.Besides,we propose the solution architecture of our strategy,which mainly includes the above roles,the network model,the user behavior prediction model and the caching strategy model.This paper does not focus on the mobility of individual users but collects the content popularity in each base station to be the characteristics of user crowd behavior.Then we adapt the classification prediction algorithm based on the historical data set of content popularity: brand new content is directly averaged,new content uses the exponential moving average(EMA)algorithm,old content uses an averaged multi-step prediction of Long shortterm memory(LSTM).With the consideration of prediction error,priority metric is introduced to revise predicted content popularity.Priority metric is the weighted average of predicted content popularity and content frequency.The weight is defined as Popularity Confidence Factor(PCF),which would adaptively adjust by the comparison of predicted content popularity and content frequency at the end of each period.APCS divides time into periods.At the beginning of each period,a part of content is prefetched into base stations according to the priority metrics.During each period,APCS would response the received request tasks,which might cause the update of caching content.At the end of each period,in addition to the PCF adjustment,historical datasets of content popularity would be updated,which might trigger a new LSTM prediction.To verify the effectiveness of our approach,this paper compares APCS with common caching algorithms(LRU,LFU,FIFO)and prefetching strategies based on content popularity(PPCS,HPCS).The simulation uses different Zipf distribution coefficients to represent different conditions of request tasks and varies in the storage capacity of base stations.The results show that APCS can achieve much better than common caching algorithms but a bit better than prefetching strategies.However,for the case of inaccurate prediction,APCS outperforms in dealing with peak traffic than prefetching strategies.