Dynamic Management Of Data Replicas Of Cloud Computing In Heterogeneous Environments

With the development and promotion of cloud computing, many applications involving mass data processing are emerging robustly. When facing with the storage requirement of mass data, there are bottlenecks in capacity and performance scalability of the traditional file system and the distributed file system becomes the basis of a variety of applications in the cloud environment. In order to enhance the system availability, reliability and fault-tolerant, the replication technology is adopted by the distributed file system.Replication technology which is a kind of data management mechanism makes multiple copies for data item putting on multiple nodes of the distributed file system respectively which is used to improve the system reliability and access efficiency. Replica management strategy mainly includes static and dynamic types. Based on the known access modes, the static replica management strategy which is suitable for the environment with steady resources determines the replica numbers and placement when the file is created. According to the changes of the resource environment, the dynamic replica management adjusts the replica numbers and placement dynamically to adapt to the changing resource environment.The replica placement strategy of the current HDFS adopting the static replica management mechanism and whose default replica factor is three is that one of the replicas is stored on a node of the local rack, while the second is putted on another node of the same rack, and the last is placed on a node of another rack. HDFS improves the data reliability and access efficiency by multiple replicas, but there are still some deficiencies of the HDFS replica management, for example, the number and the created time of the replicas are fixed and it does not consider the node heterogeneity when placing the replicas. In order to resolve these problems, a dynamic management model of data replicas of cloud computing in heterogeneous environments, DMDR, is presented. When placing the data replicas, DMDR selects the optimal nodes according to the nodes’ performance. Based on the gray prediction technology, DMDR predicts the data accessing frequency and dynamically adjusts the number of replicas according to the characters of data accessing. The experimental results show that in heterogeneous resource environments, compared with current HDFS, DMDR reduces the amount of data transferring, improves the load balance, and degrades the job finish time.