Utilization Control and Optimization of Real-Time Embedded Systems

Real-time embedded systems have been widely deployed in missioncritical applications, such as avionics mission computing, highway traffic control, remote patient monitoring, wireless communications, navigation, etc. These applications always require their real-time and embedded components to work in open and unpredictable environments, where workload is volatile and unknown. In order to guarantee the temporal correctness and avoid severe underutilization or overload, it is of vital significance to measure, control, and optimize the processor utilization adaptively. A key challenge in this mission is to meet real-time requirements even when the workload cannot be accurately characterized a priori. Traditional approaches of worst-case analysis may cause underutilization of resources, while Model Predictive Control (MPC) based approaches may suffer severe performance deterioration when large estimation errors exist. To address this challenging problem and provide better system performance, we have developed several important online adaptive optimal control approaches based on advanced control techniques. Our approaches adopt Recursive Least Square (RLS) based model identification and Linear Quadratic (LQ) optimal controllers to guarantee that the systems are neither overloaded, nor underloaded. These proposed approaches, as well as the associated tools, can quickly adapt to volatile workload changes to provide stable system performance. To minimize the impact of modeling errors, we adopt the Adaptive Critic Design (ACD) technique and develop an improved solution that requires little information of the system model. To deal with the discrete task rates, we further propose to utilize the frequency scaling technique to assist the utilization control and optimization. The computational overhead of centralized approaches explodes as the scale of systems increases. To ensure system scalability and global stability, decentralized control and optimization approaches are desired. We leverage an efficient decoupling technique and derive several distributed approaches. These approaches adopt one feedback loop to adjust the task rate, and apply another feedback loop to control the CPU frequency asynchronously. As these two manipulated variables (i.e., the Full text available at: http://dx.doi.org/10.1561/1000000042