Achieving high availability in a replicated file system by dynamically ordering transactions

A pessimistic consistency-control algorithm is presented. The principle of the algorithm is to order transactions dynamically in the global serialization graph of the system. This order is built by associating numbers with transactions in such a way that these numbers decrease along any path in the graph. Transaction numbers correspond to an arbitrary ordering of replicated objects. The model of the distributed system and the transaction manager are described. Replicated objects and the basic access strategy used to illustrate the method are introduced. Both the access algorithm, with a proof of its correctness, and the recovery algorithm are described. The generalization of the method to other pessimistic algorithms and several related issues are discussed