This paper describes an enhanced replica synchronization mechanism built in Microsoft's WinFS replica management system. The system reconciles autonomously-operating replicas in a completely peer-to-peer manner, without employing a central master or locking. The main challenge is for two replicas to exchange meta-information efficiently about (potentially numerous) data objects in order to discover what updates they are missing, and detect conflicts. The paper introduces a novel bundling mechanisms called VS, that groups together multiple objects and represents their state in a single version-vector. VS provides improved storage and communication overheads over previously known optimistic replication schemes, in the following sense. Under normal, low-fault situations, it maintains and communicates as little as a single version vector in order to represent precedence ordering of the entire set of data objects. Moreover, under settings of severe communication disruptions, VS degenerates to no worse than a single vector per object. This dramatically improves the complexities described in a preliminary write-up of the WinFS replication scheme. The VS mechanism has potentially wide applicability as a mechanism for compactly handling synchronization of arbitrarily overlapping groups of objects.

PDF file

In  ACM Operating Systems Review

Details