On Soft Real-Time Implementation of LQG Controllers

In modern embedded systems, an increasing number of controllers are implemented using soft real-time approaches, for which a controlled number of deadline misses can be tolerated. The strong push towards these approaches is mainly dictated by the need of fully exploit the available computation time or to scale down power consumption according to the operating conditions. Different models of computation have been proposed that, in combination with a predictable scheduling mechanism, produce closed loop systems with guaranteed levels of performance despite of the stochastic nature of the delay that affects the executions. One of the key issue is if output samples from the plant should be buffered during a delayed execution or simply dropped to retain the freshest. This paper explores this tradeoff for three different models of computation applied to a classic Linear Quadratic Gaussian (LQG) controller.