Event-Based Control of Stochastic Systems with Application to Server Systems

With the current strong trend towards networked and autonomous systems, it becomes less realistic to demand that all elements of a control loop should operate in a synchronous, time-triggered fashion. Above the lowest level of feedback control, it is often more natural and efficient to communicate, decide, and act based on events. Previous work shows that event-triggered control can achieve both lower average sampling rates and better performance than standard, periodic control. There is however not yet a coherent theory for the analysis and synthesis of event-based controllers.

The aim of this project is to develop theory, tools, and design methodology for event-based control of stochastic systems. The overall goals are more efficient resource usage and better performance compared to standard sampled-data control. At the same time, the methods are aimed at a wider class of control problems, including those that combine local feedback with higher-level decision-making. Such features are common in various applications such as autonomous vehicles, traffic routing, control of computing systems, supervisory plant control, and resource management in the cloud.

In the final phase of the project, in 2022 we investigated the combined identification and optimization of server systems using queueing network models. The goal was holistic load balancing, where optimal routing probabilities in a distributed microservice application were found using optimization. Applying automatic differentiation to a mean-field fluid model of the application enabled iterative grading stepping and model refitting online. The approach was evaluated on a real microservice Cloud application running Kubernetes/Istio.