| Future embedded Systems-on-Chip (SoCs) designed at nanoscale will likely consist of tens or hundreds of (potentially energy-efficient) heterogeneous cores supporting one or several dedicated applications. For such systems, the Networks-on-Chip (NoC) communication architectures have been proposed as a scalable solution which consists of a network of resources exchanging packets while running various applications concurrently. Over recent years, embedded systems have gained an enormous amount of processing power and functionality with the ultimate goal on power and performance optimization.;In this dissertation, with the ultimate goal that any system optimization is to satisfy the end user, we study outstanding problems on embedded system methodology, while incorporating the user behavior information into the modeling, analysis, optimization, and evaluation steps. Our specific contributions are as follows. (1) For predictable system configurations derived from use-case applications, we explore the design space of system interconnect on application-specific multi-processor systems-on-chips (MPSoCs). With the proposed analytical and simulation models, we can theoretically generate fabric solutions with optimal cost-performance trade-offs, while considering various design constrains, such as power, area, and wirelength. (2) For unpredictable system configurations incorporating users interaction with the system, we present a new design methodology for automatically generating regular NoC platforms, while including explicitly the information about the user experience into the design process. Such off-line design flow aims at minimizing the workload variance and allows the system to better adapt to different types of uses. (3) For applications entering and leaving the system dynamically, we propose an efficient technique for run-time application mapping onto heterogeneous NoC platforms with the goal of minimizing the communication energy consumption, while still providing performance guarantees. The proposed technique allows for new applications to be easily added to the system platform with minimal inter-processor communication overhead. (4) To address the problem of runtime resource management in NoC platforms while considering permanent, transient, and intermittent failures, we propose a system-level fault-tolerant approach that investigates several metrics for network contention and system fragmentation, as well as their impacts on system performance. (5) Finally, having generated system platforms which exhibit less variation among the users behavior, we explore flexible and extensible run-time resource management techniques that allow system to adapt to run-time stimuli specific to each class of user behaviors; these techniques change dynamically according to the user models built on-line based on different user needs. |
