| Increasingly, the Internet is becoming a platform for distributed computing with upwards of 70 million machines and countless available compute cycles. Harnessing these cycles has been the quest for distributed computing research for the last 20 years. Recently, large metacomputing systems are being proposed and constructed that span many research organizations connected by wide-area networks. One of the challenges in building such systems is fault tolerance. As the number of machines increases, failures will become more common. In addition, with the use of wide-area networks, there is greater chance of network disruptions causing failures.; The research presented here develops a novel fault tolerance mechanism that is designed for large-scale, wide-area distributed systems. This technique, called lazy active replication, is a form of dynamic replication where objects are replicated only after a failure occurs. Care must be taken when dynamically replicating these objects to ensure that consistency is maintained. Naturally, in order to recover an object consistently, there must be recovery information relating to that object available at the time the object is recovered. So the first challenge is to build a mechanism that can distributed recovery information in the wide-area network so as to guarantee the consistency of the objects. The second challenge involves partitioned operation. Since replicas can be created due to network disruptions, there will in general by multiple copies of the same object each whose state is diverging from the other. In such an environment, the system needs to ensure that some degree of consistency is preserved.; The contributions of this dissertation include: the development of a formal consistency criteria that is appropriate for partitionable, dynamic replication systems, investigation of mechanisms to distribute recovery information in a scalable way, and the development of recovery protocols for implementing the dynamic replication itself. |
