An Adaptive Closed-Loop Approach For Timely Data Services

In data-intensive soft real-time applications, e.g., e-commerce, traffic control, and target tracking, a database system needs to process transactions in a timely manner. However, user transactions may suffer from unpredictable large delays when the database system is overloaded due to flash transaction arrivals and transaction aborts/restarts. To address the problem, we design a new adaptive closed-loop method considering database semantics to control the response time to be below a target set-point even when dynamic workloads are given. Our approach continues to update the database model at runtime and re-tunes the response time controller based on the adjusted model, because the relation between the workload and response time may vary in time. Notably, our adaptive control scheme is different from most existing closed-loop methods for real-time database performance management that model the database system and design and tune the controller entirely offline with no online adaptation. The results of the performance evaluation undertaken in a real-time database testbed show that our approach maintains the database response time below the target set-point for most of the time even under steep workload surges, quickly canceling any transient delay overshoot that exceeds the set-point. However, the tested state-of-the-art baselines fail to do it.