Using application specific knowledge for configuring object replicas

Abstract

In distributed systems, object replication is used to improve the availability and performance of applications in the presence of failures. When determining the configuration of a replicated object (i.e., number and location of replicas), a number of often conflicting factors need to be considered, e.g., the availability and performance requirements of the object. However, application specific knowledge about the objects, such as any inter-dependencies, is typically not accounted for. In many applications this information can effect an object's availability. Distributed systems which support replication typically give users only basic control over the configuration of a replicated object, such as the number or location of individual replicas. Expressing application specific knowledge is not possible. We have designed and implemented a replication sub-system allowing applications to control these aspects of replication. This system allows the efficient replication of an arbitrary number of objects with arbitrary inter-dependencies.