Power efficiency has emerged as the overarching design constraint for modern hybrid-core computing architectures. This research advances new models of the physical time, energy, and power needed to execute an algorithm. Such models aim to address several contemporary research questions in high-performance computing (HPC). For instance, if there is a limited amount of power available to a computer system, how should that power be allocated among various system components to most quickly or most energy-efficiently execute a given computation? Or, does anything need to change in the design of algorithms, given a user's or a system's explicit power or energy constraints?
The technical approach extends a preliminary model, referred to as the energy roofline, which the Principal Investigator (PI) and his team have developed as part of a prior project. This approach starts from first principles of algorithmic analysis that express the intrinsic concurrency and communication properties of an algorithm; and from that, it derives models of time, energy, and power using cost models informed directly by the behavior of real algorithms, software, and systems. The project considers several critical enhancements of the preliminary work, which is leading to a family of models of increasing accuracy and complexity. The research will have a broad impact on a wide range of important design issues in exascale systems.
