Abstract: Cloud storage is getting more and more popular as a new trend of data management. Data replication has been widely used as a means of increasing the data availability in large-scale cloud storage systems where failures are normal. However, most data replication schemes do not fully consider cost and latency issues when users need large amounts of remote replicas. We present animproved dynamic replication management scheme (IDRMS). By adding a prediction model, the optimal allocation of replicas among the cloud storage nodes is determined that the total communication cost and network delay are minimal. When the local data block is frequently requested, the data replicas can be moved to a closer or cheaper node for cost reduction and increased efficiency. Moreover, we replace
the B+ tree with the B*tree to speed up the search speed and reduce workload with the lowest blocking probability. We define the value of popularity to adjust the placement of replicas dynamically. We divide the data nodes in the network into hot nodes and cool nodes. By changing to visit cool nodes instead of hot nodes, we can balance the workload in the network. Finally, we implement IDRMS in Matlab simulation platform and simulation results demonstrate that IDRMS outperforms other replication management schemes in terms of communication cost and load balancing for large-scale cloud storage.

Key words: cloud storage, dynamic replication management, blocking probability, prediction, load balancing