Modern architecture present in self-power devices such as mobiles or tablet computers proposes the use of asymmetric processors that allow either energy-efficient or performant computation on the same SoC.

For energy efficiency and performance consideration, the asymmetry resides in differences in CPU micro-architecture design and results in diverging raw computing capability. Other components such as the processor memory subsystem also show differences resulting in different memory transaction timing. Moreover, based on a bus-snoop protocol, cache coherency between processors comes with a peculiarity in memory latency depending on the processors operating frequencies.

All these differences come with challenging decisions on both application schedulability and processor operating frequencies. In addition, because of the small form factor of such embedded systems, these devices generally cannot afford active cooling systems. Therefore thermal mitigation relies on dynamic software solutions.

Current operating systems for embedded systems such as Linux or Android do not consider all these particularities. As such, they often fail to satisfy user expectations of a powerful device with long battery life.

To remedy this situation, this thesis proposes a unified approach to deliver high-performance and energy-efficiency computation in each of its flavours, considering the memory subsystem and all computation units available in the system. Performance is maximized even when the device is under heavy thermal constraints. The proposed unified solution is based on accurate models targeting both performance and thermal behaviour and resides at the operating systems kernel level to manage all running applications in a global manner.

Particularly, the performance model considers both the computation part and also the memory subsystem of symmetric or asymmetric processors present in embedded devices. The thermal model relies on the accurate physical thermal properties of the device. Using these models, application schedulability and processor frequency scaling decisions to either maximize performance or energy efficiency within a thermal budget are extensively studied.

To cover a large range of application behaviour, both models are built and designed using a generative workload that considers fine-grain details of the underlying microarchitecture of the SoC. Therefore, this approach can be derived and applied to multiple devices with little effort.

Extended evaluation on real-world benchmarks for high performance and general computing, as well as common applications targeting the mobile and tablet market, show the accuracy and completeness of models used in this unified approach to deliver high performance and energy efficiency under high thermal constraints for embedded devices.