Resource and performance trade-offs in real-time embedded control systems

Resumen

The use of computer controlled systems has increased dramatically in our daily life. Microprocessors are embedded in most of the daily- used devices. Due to cost constraints, many of these devices that run control applications are designed under processing power, space, weight, and energy constraints, i.e., with limited resources. Moreover, the embedded control systems market demands new capabilities to these devices or improvements in the existing ones without increasing the resource demands. Enabling devices with real-time technology is a promising step toward achieving cost-effective embedded control systems. Recent results of real-time systems theory provide methods and policies for an efficient use of the computational resources. At the same time, control systems theory is starting to offer controllers with varying computational load. By combining both disciplines, it is theoretically feasible to design resource-constrained embedded control systems capable of trading-off control performance and resource utilization. This thesis focuses on the practical feasibility of this new generation of embedded control systems. To this extend, two issues are addressed: 1) the effective implementation of control loops using real-time technology and 2) the evaluation of resource/performance- aware policies that can be applied to a set of control loops that concurrently execute on a microprocessor. A control task generally consists of three main activities: input, control algorithm computation, and output. The timing of the input and output actions is critical to the performance of the controller. The implementation of these operations can be conducted within the real- time task body or using hardware functions. The former introduces considerable amounts of jitters while the latter forces delays. This thesis presents a novel task model as a computational abstraction for implementing control loops that is shown to remove the endemic problems caused by jitters and delays. This model is synchronized at the output instants rather than at the input instants. This has been shown to provide interesting properties. From the scheduling point of view, the new task model can be seamlessly integrated into existing scheduling theory and practice, while improving task set schedulability. From a control perspective, the task model absorbs jitters because it allows irregular sampling by incorporating predictors, and improves reactiveness in front of perturbations. In addition, Kalman techniques have been also investigated to deal with the case of noisy measurements. The effective implementation of simple control algorithms making use of this new task model does not guarantee the feasibility of implementing state-of-the-art resource/performance-aware policies. These policies, which can be roughly divided into feedback scheduling and event-driven control, have been mainly treated from a theoretical point of view while practical aspects have been omitted. Conversely to the initial problem targeted by these policies, that is, to minimize or keep resource requirements to meet the tight cost constraints related with mass production and strong industrial competition, research advances seem to require sophisticated procedures that may impair a cost-effective implementation. This thesis presents a performance evaluation framework that permits to assess these policies in terms of the potential benefits offered by the theory as well as the pay-off in terms of complexity and overhead. The framework design is the result of a taxonomical analysis of the related state-of-the-art. Among other specifications, the framework, which is composed by a simulation and an experimental platform, supports both event/time triggered paradigms, allows different sort of control and optimization algorithms, and flexibly evaluates control performance and resource utilization.