Towards an Optimized Cloud Replication Protocol

摘要

The problem of efficient data replication across data centers is known to be challenging due to the amount of data that needs to be communicated, the load on the network and individual data centers to execute such replication, and the inherent penalties for not generating replicas in a timely fashion, e.g., loss of data due to failures in the source data center. Optimizing only for network costs is equivalent to solving the problem of minimum cost multicast. However, overall cost is not the best metric to (i) increase reliability of data that is being replicated, or (ii) balance the costs of individual data centers in the replication process. This paper characterizes optimal strategies for a sequential replication process from a mathematical model and evaluates performance considering different transport protocols and their sensitivity to round-trip-time (RTT) and packet losses. This paper also implements and deploys four key strategies, including the optimal sequential solution, to thoroughly evaluate performance of the implementation for different file sizes. We show that the optimal strategy can reduce transmission time by 37 % in average with respect to the best alternative strategy. We also show that random selection strategies can incur in large performance penalties even in simple deployments.