Highly available atomic objects

Abstract

The author describes an approach to implement highly available atomic objects in a distributed system. The approach uses the semantics of an object's type to increase availability and concurrency. The increase comes at the expense of writing a layer of type-specific code for concurrency control. The type-specific code can be written without making any assumption of any underlying replication or the serialization order used, allowing it to be reused more easily. He uses the virtual partition algorithm (A. Al Abbadi et al., 1986) for replica control. It avoids reading multiple replicas during normal operation at the expense of synchronizing the replicas when they re-establish communication. The optimization is important in the context of semantics-based implementations because potentially no replica is more up-to-date than the others. In contrast to similar semantics-based approaches, this approach assigns quorums to operations of the nonatomic objects used to implement the atomic objects, rather than to operations of the atomic objects.<>