Concurrency and availability as dual properties of replicated atomic data

Abstract

A replicated object is a typed object that provides a set of operations to its clients. A quorum for an operation is any set of sites whose co-operation suffices to execute that operation. An operation's quorums determine its availability, and constraints on quorum assignment determine the range of availability properties realizable by replication. This paper compares the constraints on availability and concurrency imposed by three classes of atomicity mechanisms, respectively encompassing generalized two-phase locking, multiversion timestamping, and hybrid techniques. An analysis of the constraints on quorum assignment necessary to maximize the concurrency permitted within each class shows that only the class encompassing hybrid mechanisms is undominated: (1) Hybrid schemes permit more quorum assignments than timestamping schemes, even though they place incomparable constraints on concurrency. (2) Hybrid and locking schemes place incomparable constraints on quorum assignment, even though hybrid schemes support more concurrency. (3) Timestamping and locking schemes are incomparable with respect to both concurrency and quorum assignment. (4) Bounding the maximum depth of transaction nesting tightens constraints on concurrency for ail three classes, but reduces the constraints on quorum assignment for hybrid schemes only. (5) Bounding the number of versions retained by multiversion timestamping schemes reduces concurrency but enhances availability.

Replication—Support for Advanced Mobile Applications

“…A support system has to know what kind of replication-support the application requires. Our approach is to use a description of the objects' semantics via a seriaI-dependency-relation (Herlihy, 1990) and a lean classification of application-specific consistency requirements. Infonnally, a serial-dependency-relation tells what subset of operations executed on all replicas has to be known when one wants to be sure that another operation returns a consistent value.…”

Roles of Resource and Data Contention on the Performance of Replicated Distributed Database Systems

“…In a fully replicated database system, data are redundantly stored at each node. System reliability is enhanced at the expense of storage and update cost (Carey & Livny, 1991;Herilhy, 1990). For such systems, the consistency problem is more complex.…”