Time-Based Coordinated Checkpointing

Distributed systems are being used to support the execution of applications ranging from long-running scientific simulators to e-commerce on the Internet. In this type of environment, the failure of one of its components, either a computer or the network, may prevent other components from completing their tasks. Since the probability of failure increases with the number of computers and execution time, it is likely that these applications will be interrupted unless provision is made for failure handling. In this thesis we address the problem of fault recovery in distributed systems. The thesis describes two variations of a coordinated checkpoint protocol that uses time to remove most causes of overhead, and to avoid all types of direct coordination. The time-based protocol does not have to transmit extra messages, does not need to tag the application messages, and only accesses the stable storage when the checkpoints are saved. The thesis also describes a new coordinated checkpoint protocol that is well adapted to mobile environments. It uses time to indirectly coordinate the creation of new global states, and it saves two different types of checkpoints to adapt its behavior to the current network characteristics. Traditional techniques for fault diagnosis in distributed systems, either based on watch-dogs or polling, exchange performance with detection latency. The thesis introduces a complementary mechanism that uses the error codes returned by the stream sockets. Since these errors are generated automatically when there is communication with a failed process, the mechanism incurs only in small overheads. Our results show that, in most cases, failures could be located using only the errors from the sockets. A large number of checkpoint-based recovery protocols have been proposed in the literature, however, most of them were never evaluated. The thesis describes the design and implementation of a run-time system for clusters of workstations that allows the rapid testing of checkpoint protocols with standard benchmarks. {\it RENEW - Recoverable Network of Workstations} provides a flexible set of operations that facilitates the integration of checkpoint and rollback recovery protocols.