Modeling Energy Consumption of Computing Systems: from Homogeneous to Heterogeneous Systems

Nowadays, reducing energy consumption and improving the energy efficiency of computing systems become ones of the leading research topics in computer science. In order to improve energy efficiency, it is crucial to understand how computing systems consume energy. Power and energy models provide prediction and insight into how computing systems consume power and energy. However, it is challenging to build accurate power and energy models that can be used for general devices and general applications. In this thesis work, we propose three approaches of devising power and energy models, varying from homogeneous systems including one type of devices (e.g., CPU, GPU, ultra-low-power embedded system) to heterogeneous systems including several types of devices with different architectures. We developed new fine-grained power models supporting architecture-application co-design by considering both platform and application properties. The models were trained and validated with data from a set of micro-benchmarks and application kernels on Movidius Myriad, an ultra-low power embedded system. We also proposed and validated a framework predicting when to apply race-to-halt strategy to a given application. We devised ICE, a new energy complexity model for parallel (multi-threaded) algorithms that were validated on real multicore platforms and applicable to a wide range of parallel algorithms. The study also provided the platform parameters of the ICE models for eleven platforms including HPC, accelerator and embedded platforms to improve the model usability and accuracy. We proposed REOH, the holistic tuning approach to choose the most energy-efficient configurations for heterogeneous systems including several types of devices with different architectures (e.g., CPUs, GPUs). Based on the REOH approach, we developed an open-source energy-optimizing runtime framework for selecting an energy efficient configuration of a heterogeneous system for a given application at runtime