1 Department of Applied Mathematics and Computer Science, Technical University of Denmark2 Embedded Systems Engineering, Department of Applied Mathematics and Computer Science, Technical University of Denmark3 Copenhagen Center for Health Technology, Center, Technical University of Denmark
Embedded systems are becoming increasingly complex and have tight competing constraints in terms of performance, cost, energy consumption, dependability, flexibility, security, etc. The objective of this thesis is to propose design methods and tools for supporting the tradeoff analysis of competing design objectives during the early design phases, which are characterized by uncertainties. We consider safety-critical real-time applications modeled as task graphs, to be implemented on distributed heterogeneous architectures consisting of processing elements (PEs), interconnected by a shared communication channel. Tasks are scheduled using fixed-priority preemptive scheduling, and we use non-preemptive scheduling for messages. As a first step, we address the problem of function-to-task decomposition. In this context we have assumed that the application functionality is captured by a set of functional blocks, with different safety requirements. We propose a Genetic Algorithm-based metaheuristic to solve the function-to-task decomposition problem. Our algorithm also decides the mapping of tasks to the PEs of a distributed architecture and the reliability of each PE in the architecture, such that the safety and integrity constraints are satisfied, the schedulability of the real-time application is guaranteed and the overall development and product unit costs are minimized. Next, we investigate tradeoffs between performance, energy and reliability. Addressing energy and reliability simultaneously is especially challenging, since lowering the voltage to reduce the energy consumption has been shown to increase the transient fault rate. We are interested to tolerate transient faults and we use task replication for recovery. We propose a Tabu Search-based approach, which decides the mapping of tasks to processing elements, as well as the processor voltage and frequency levels for executing each task, such that transient faults are tolerated, the real-time constraints of the application are satisfied, and the energy consumed is minimized. In this thesis, we target the early design phases, when decisions have a high impact on the subsequent implementation choices. However, due to a lack of information, the early design phases are characterized by uncertainties, e.g., in the worst-case execution times (WCETs), in the functionality requirements, or in the hardware component costs. In this context, we select the hardware components for the architecture and derive a mapping of tasks in the application, such that the resulted implementation is both robust and flexible. The architecture also has a high chance to have its unit cost within the cost budget. Robust means that the application has a high chance of being schedulable, considering the WCET uncertainties, whereas a flexible mapping has a high chance to successfully accommodate future functionality changes. We propose a Genetic Algorithm-based approach to solve this optimization problem. The proposed tradeoff analysis methods have been evaluated using several synthetic and real-life benchmarks.