Replication is often used in distributed hash tables (DHTs) to reduce the latency of query requests. For read-write objects, replicas are commonly kept up-to-date by having a master node maintain physical pointers to replica hosts. Since these links are based on IP addresses, they must be updated whenever a linked node fails. This can be costly in systems with high churn. In this paper, we introduce a replication protocol whose replica links employ both physical and logical node identifiers. The resulting logical links are based on the DHT identifier and can be used to correct the physical links if they are wrong; as a result, the correctness of the physical links can be relaxed. This protocol significantly reduces the cost of maintaining physical links. We also propose a novel cost model, which is used to determine when and where replicas should be created. Our cost model calculates the benefit of a replica by both the number of queries a node receives and the distance between the node and the query requester nodes.
展开▼
