Towards co-designed energy efficient computing and runtimes: simulation-framework and experiments
Date
2015
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
As silicon field effect transistor based computing continues into the twenty-first century, there has been a shift in the design of computational hardware systems. Having run solidly into several walls including the power and memory walls, computer architects are now taking different tactics to translate Moore’s law into more performance. Whether by using more leakage-aware transistor technology, providing multiple ISA equivalent microarchitectures to match variable computing demands, or including fixed function accelerators, engineers continue to push the limits of physics and mass production techniques to deliver better computing systems. However, this dissertation will provide evidence for an equally important need to shift system software and its organization to both leverage the new hardware capabilities and organization. To that end, the proposal will present 3 case studies: the current ability for system software to manage a processor’s power envelope through hardware mechanisms using a tool called REST as a prototype, multiple novel chip level approximations required to model heterogeneous near threshold many-core systems of the future, and lastly, an initial runtime driven algorithm that performs dynamic tiling and scheduling based on a selfaware runtime. All these pieces combined will guide architects and system software developers in creating a new system of computation that can meet the energy efficiency goals of an exascale machine.