Multiclass replicated data management: exploiting replication to improve efficiency

Research efforts in replication-control protocols primarily use replication as a means of increasing availability in distributed systems. It is well-known, however, that replication can reduce the costs of accessing remotely-stored data in distributed systems. We contribute a classification of replicas and a replication-control protocol which introduce the availability benefits of replication and, at the same time, exploit replication to improve performance, by reducing response time. Each replica class has different consistency requirements. Metareplicas keep track of up-to-date replicas for recently-accessed objects and help exploit data-reference localities. Thus they allow many transaction operations to execute synchronously at only a single (and often local) replica. Pseudoreplicas are nonpermanent replicas that facilitate “localized execution” of transaction operations. True replicas are ordinary, permanent replicas as used in other replication schemes. For many commonly occurring replication scenarios, the protocol outperforms both replication-control protocols in the literature and nonreplicated systems, while offering the availability benefits of replication