A Fault-Tolerant Strategy of Redeploying the Lost Replicas in Cloud
Ning Wang, Yang Yang, Zhenqiang Mi, Qing Ji, Kun Meng · 2014
In cloud storage centers, replica of file may be lost subjected to the failure of nodes, which will affect the efficiency of file access, as well as users' satisfaction. To cope with this problem, the method of redeploying the lost replicas on some other servers to maintain system availability is often adopted. Normally, a file is divided into many blocks with the same size and the popularity of the blocks is different in cloud storage system, which could be used as a parameter in deploying replicas. Therefore, in this paper, Scarlett system is utilized to determine the optimal number of block replica based on the block popularity. Then, considering the system load, the total cost and quality of services, we present a selective data recovery method subjected to the failure of nodes. In the meantime, a cost-efficient replicas deployment strategy, namely CERD, is provided. The strategy has been verified in HDFS. Finally, we simulate the environment with random cloud node failure, and compare our strategy with the strategies of Hadoop default. The results verify that CERD strategy can balance the load of the whole system, reduce the total cost of service, as well as provide higher service quality, which are consistent with the theoretical analysis.