Peer Replication with Selective Control

Peer-to-peer optimistic replication strategies provide improved functionality over traditional client-server models by enabling any-to-any communication. This is especially useful in environments with common network partitions, such as mobile environments, in which communicating with a closely connected peer can be cheaper than communicating with a distant, weakly connected server. However, most peer solutions require full replication, meaning that all replicas store the entire replication unit, whether it be a volume or a database. Such strategies are clearly ineecient and not cost eeective. Users are forced to both store more data than they require and spend time and money maintaining consistency on that data, potentially making peer solutions unusable in practice. We have developed a set of algorithms that implement selective replication, allowing individual portions of the replication unit to be replicated independently. We have implemented our solution in Ficus and allow les from the volume to be selectively replicated without storing the entire volume. We present a description of the algorithms and their implementation in the Ficus le system, as well as a performance analysis. We argue that these methods permit the practical use of peer optimistic replication.