Data replica management is an important part of cloud computing system. When processing massive data in a cloud computing system, the existing algorithms for data storage and resource scheduling do not consider the dynamics and reliability of data replicas．We therefore propose a dynamic replica placement mechanism based on the domain structure. It takes into full account the number and position of data replicas, as well as the  cost of system resources such as memory and bandwidth when a replica is produced. Firstly, according to the information of the replica, the data with high access frequency and long average response time, are replicated, and how to calculate the number of replicas is explained. Secondly, in order to reduce the selection range of nodes of the replica distribution, we propose a dynamic replica placement algorithm, which can choose the range of placement for the replicas according to domain division. Experimental results show that the proposed algorithm can significantly reduce the waste of storage space for the replica with low access frequency, as well as the transmission delay across nodes for the replicas with high access frequency. Besides, it effectively improves the access efficiency to data files in cloud storage systems, the load balance, and the reliability and availability of cloud storage systems.