Research On Efficient Caching Policy Based On Machine Learning

With the rapid development of the mobile Internet industry,image service providers are facing the serious challenges brought by the storage,transmission and processing of massive images.An excellent cache policy can not only reduce the user access time,but also improve the performance of the cache system.Previous researchers used machine learning algorithms to develop intelligent policies such as cache prefetching,access and replacement policies.However,there are two common problems with current cache policies.Firstly,they do not consider the cost of misclassifying objects in the context of actual cache scenarios.Secondly,the characteristics of data access pattern are not fully utilized to optimize the configuration management of cache space,and the cache resources of the system are not maximized.For the first problem,we propose a size-aware cost-sensitive classification algorithm,SAda Cost,to predict whether photos are re-accessed.To improve the prediction accuracy as much as possible,SAda Cost introduces two cost parameters and considers the cache overhead caused by missclassifying the re-accessed objects and missclassifying the larger objects that will not be re-accessed.For the second problem,we propose an adaptive segmented caching policy,AS3 LRU,which designs the promote rules according to access frequency,and adjusts the allocation ratio of each segment based on the cache hit ratio within a certain time window.Based on the two algorithms,we propose an efficient cache policy,i.e.,SAda Cost-AS3 LRU,which can utilize the cache space more efficiently to improve the hit rate and byte hit rate.It firstly predicts whether objects will be re-accessed,then puts the objects into the different cache segments according to the prediction results.We conduct the experiments using QQPhoto from Tencent Inc.,the largest social network service company in China.Experimental results show the proposed approach outperforms in cache hit rate,cache byte hit rate and response time.Compare with LRU which is used by Tencent,our approach improves the hit rate by 19.10%,byte hit rate by21.68%,and reduce the average response time by 9.89%.Compare with S3 LRU,our approach improves the hit rate by 7.77%,byte hit rate by 9.42%,and reduce the average response time by 4.85%.