Adaptive protocols for survivability of transactions operating on replicated objects

We present a novel transaction model for distributed safety-critical real-time applications in which transactions become increasingly critical under failures of subsequent incarnations. When a transaction-dependent criticality threshold has been reached, the transaction is considered essentially critical or hard, i.e. the deadline has to be met lest the system is doomed to be in a disastrous state. The corresponding adaptive measures (criticality, sensitivity) to provide that potentially all hard deadlines are met, i.e. that the system survives, are based on the adaptive services of the distributed real-time operating system MELODY. Transaction criticality and sensitivity are stepwise defined starting from the corresponding measures for the tasks constituting the transaction. The authors propose adaptive techniques for distributed concurrency control under safety-critical real-time requirements such as evolving from the modification and implementation of standard concurrency control protocols. In particular, we refrain from traditional features like roll-back, or restart after preemption, because they are costly and have an unpredictable effect on meeting hard deadlines. Through extensive distributed experiments we compare the performance of a recent concurrency control algorithm (O2PL) to its adaptive version O2PL.ADT, resulting in a clear advantage of the latter, both in terms of deadline failure rates and survivability. Since the MELODY model has been extended by adding reactive features, we confirm our findings in experiments with the reactive protocol versions O2PL.SIM and O2PL.ADT.SIM.