The goal of this project is to lay out strategies and abstractions for developing lower-level system modeling and parameter tuning techniques, and for linking these to higher-level architectural and system management approaches. The project also aims to back existing post-manufacturing tuning techniques with deeper analytic methods, and to augment these one-time adjustments with more dynamic management techniques that are ongoing as the system is in use. To make the shift from very static (one-time) tuning towards very dynamic (ongoing) power-performance tuning, one needs models whose detail and abstraction vary. This allows fast, but abstract models to be used dynamically by the operating system, in order to dynamically adjust power/thermal behavior in order to stay within budget while also meeting performance goals. Microarchitectural techniques can also be employed, but likewise need to be informed by good models and measurement techniques to guide their use. The project will address these modeling issues and the associated on-the-fly management techniques.
The research program pursues broad impact in several ways. First, the project has an important component for knowledge dissemination and technology transfer. The modeling and management techniques proposed in this project will be disseminated and released for free use. Building on a record of strong support for undergraduate research and underrepresented groups, the PIs will continue and broaden such activities through this collaboration. Because of the geographic proximity of CMU and Princeton, group meetings unifying the two efforts will be possible throughout the project.
