Context-specific synchronization for atomic data types

Abstract

Highly concurrent and reliable atomic data types are crucial for the design of distributed systems. Deferred update (DU) and update-in-place (UIP) are two common recovery strategies for implementing atomic data types. These two strategies place incomparable constraints on the conflict relations between concurrent operations resulting in incomparable synchronization protocols. Also, the conflict relations used are usually static in the sense that the algorithms do not use the context-specific information that may be available in the system. In this paper, a new synchronization mechanism that employs a hybrid recovery scheme by using both DU and UIP is proposed. Furthermore, the protocol is dynamic in the sense that the context-specific information is also used to determine conflict relations among concurrent operations on the atomic data types.Another extension is the use of ordered shared relationship between locks to execute conflicting operations concurrently. Thus, the execution of operations is never delayed in the proposed protocol, however, the commitment of the transactions invoking these operations may be delayed due to the restriction imposed by the ordered shared relationship between locks. We also demonstrate that the sets of histories accepted by a two phase locking protocol using either DU or UIP are proper subsets of the set of histories accepted by the proposed protocol.

This research is supported by the NSF under the grant numbers IRI-9004998 and IRI-9117094.