Research On Big Data Placement For Low Latency Based On Refinforcement Learning

In the era of BigData,low latency is a major problem for big data enterprises and companies.The reduction of latency not only improves the processing performance of data center,but also improves the service quality of users At present,most of the exsiting methods to reduce the latency by adopting heuristic algorithm or static models.However,traditional methods are difficult to adapt to the dynamic data center environment.From the perspective of data placement in distributed system,each data placement can affects data read/write latency,in this paper,two algorithms are proposed which are K-means and Deep Q-Learning DQL.According to the feature of the network topology and the regular of data read/write in data center.First,in view of the topology structure of the data center network and the law of data reading and writing,we adopt clustering algorithm K-means to cluster the storage nodes and then sort the center of clustering.The nodes close to the center of clustering are employed.Second,for the dynamic user requests and network configuration,we take data placement into markov decision process(MDP)process,the reinforcement Learning(Deep Q-Learning,DQL)aims at fitting QL state function and optimal data placement strategy.In our design,data read/writes requests learn and acquire reward from data center environment.The main contents of this paper are as follows:First,we analyze the data center network topology characteristics and data read/write requests.We adopt K-means algorithm and build the network environment of mininet to simulate data center.and experimental simulation was carried out with the algorithm to Through experimental comparison,K-means algorithm reduces data read-write latency by 61.5%and 53%,respectively,compared with traditional algorithm CommonIP(the data is as close as possible to the IP address that frequently accesses it).Second,under the scenario of data deployment with dynamic changes,we adopt the Deep Q-Learning algorithm,For the problem of large input space of q-learning state,we use DQL deep neural network to fit the value function.In this paper,DQL state space is further sparse to reduce input of neural network,so as to improve performance.We adopted the same experimental environment as K-means algorithm,and through our simulation comparison,DQL reduced data read and write latency by 63.9%and 82.4%,respectively,compared with CommonlP.Experimental results show that the DQL algorithm has strong applicability in dynamic data center network environment.