The Replica Management System: a scheme for flexible and dynamic replication

The actual gains achieved by replication are a complex function of the number of replicas, the placement of those replicas, the replication protocol, the nature of the transactions performed on the replicas, and the availability and performance characteristics of the machines and networks composing the system. This paper describes the design and implementation of the Replica Management System, which allows a programmer to specify the quality of service required for replica groups in terms of availability and performance. From the quality of service specification, information about the replication protocol to be used, and data about the characteristics of the underlying distributed system, the RMS computes an initial placement and replication level. As machines and communications systems are detected to have failed or recovered, or performance characteristics change, the RMS can be re-invoked to compute an updated mapping of replicas which preserves the desired quality of service. The result is a flexible, dynamic and dependable replication system