A quantitative approach to run-time monitoring of Markovian concurrent systems

Techniques for modeling and evaluation of timed concurrent systems can be effectively employed to support runtime monitoring of safety-critical systems. We address the issue by developing a predictive analysis engine that can guide a distributed sensor network to adaptively monitor a system that depends on a set of components, so as to guarantee reliable observation of critical system conditions while restraining the number of observation samples. The monitored system is abstracted by a Fault Tree (FT) that represents the combinations of components conditions that lead to a critical system event. In turn, components are assumed to have recurrent behavior with exponential sojourn time in each operation mode and they are specified by Continuous Time Markov Chains (CTMCs). On the one hand, a predictive analysis engine based on the fault-tolerance architecture receives observations of components from the monitoring utility and maintains a dynamic measure of the transient probability that the overall system is in a critical condition. On the other hand, the monitoring utility receives from the analysis engine a measure of the contribution of each component to each predicted critical condition of the system. This enables the definition of sampling strategies that adapt the sample rate of nodes to the actual behavior of the supervised system. Effectiveness of adaptive sampling with respect to fixed-rate sampling is experimented in a setting where the analysis engine is integrated with a monitoring utility through an Enterprise Service Bus (ESB).