A First-Principles Approach to Performance, Power, and Energy Models for Contemporary Multi- and Many-Core Processors

Models are an integral part of any engineering process and therefore an essential tool in scientific and high-performance computing: In particular, performance models assess the efficiency of implementations and, if needed, help locate bottlenecks and select suitable optimization approaches. In the same way, power and energy models are used to meet similar demands put forth from within the entire spectrum of computing: from the increase of battery life in mobile devices to sustainable power budgets for next-generation supercomputers. The work carried out in this thesis can be summarized as two main efforts, the first of which is the development of performance, power, and energy models for loop-based streaming codes. The proposed models are built on first principles and specifications but rely on selected empirical data to increase their precision; as a result, the presented models deliver estimates of unprecedented accuracy. Generality as a design goal is addressed by taking different processors from various vendors—including Intel, AMD, and IBM—into account. The second effort of this thesis is to improve the community's understanding of performance-, power-, and energy-related behavior of contemporary server processors by deriving insight from the models' first principles. In this respect, the scientific contributions include the identification of universal behavior, the uncovering of governing mechanisms, and the creation of best practices to optimize performance as well as power and energy consumption.