A Study of Roll–Forward Recovery from Faults among Communicating Processes

Roll-Forward Checkpointing Schemes combine fault-detectionusing duplicated processes and checkpoints with a characteristic fault-treatment strategy. In existing schemes, faults of single processes are detected by checkpoint-comparison, localized and eliminated by a retry-execution on a spare processor and lookahead execution in parallel. The main advantage of these schemes is a completion time in occurrence of faults close to the time of a faultless run. Only in improbable situations a rollback is necessary. This paper applies these schemes for communicating user processes and proposes an extension with reaction schemes of communicationdependent processes. The resolution of problemswhich arise due to communicating processes improves the strategy, because communication provides an additional possibility to detect and localize faults. Beside the specification of a fault-detecting protocol for the communication of duplex systems, several variants of Roll-Forward strategies for communicating and interacting processes are proposed. Variants originate by the applied treatment scheme of the faulty duplex system and the reaction scheme of distant processes. Two extreme behavioral cases can be distinguished. One variant treats faults with a slower progress of the application but reaches after the treatment a duplex-certified state immediatly. The other extreme is, that the progress of processes is influenced only insignificantly but the state of the processes is lacking the duplex-certification for a longer time period. The duration from the fault occurrence to the reentry in a duplex certified fault free state may take a longer time than it would be lost using the first variant. The choice of variants is important when interactions are necessary during processing and these interactions have to result from a duplex certified fault free state but also have to meet deadlines.