Value-driven multi-class overload management in real-time database systems

Abstract: In complex real-time applications, real-time systems handle significant amounts of information that must be managed efficiently, motivating the need for incorporating real- time database management into real-time systems. However, resource management in real- time database systems is a complex issue. Since these systems often operate in environments of imminent and transient overloads, efficient overload handling is crucial to the performance of a real-time database system. In this thesis, we focus on dynamic overload management in real-time database systems. The multi-class workload consists of transaction classes having critical transactions with contingency transactions and non-critical transactions. Non-critical transaction classes may have additional requirements specifying the minimum acceptable completion ratios that should be met in order to maintain system correctness. We propose a framework which has been implemented and evaluated for resolving transient overloads in such workloads. The contributions of our work are fourfold as the framework consists of (i) a new scheduling architecture and (ii) a strategy for resolving transient overloads by re-allocating resources, (iii) a value-driven overload management algorithm (OR-ULD) that supports the strategy, running in 0{n log n) time (where n is the number of transactions), and (iv) a bias control mechanism (OR-ULD/BC). The performance of OR-ULD and OR-ULD/BC is evaluated by extensive simulations. Results show that, within a specified operational envelope, OR-ULD enforces critical time constraints for multi-class transaction workloads and OR-ULD/BC further enforces minimum class completion ratio requirements.