The Institute for Sustained Performance, Energy, and Resilience
- University of California, San Diego, CA (United States)
The Green Queue framework is a framework designed to automate the development and deployment of customized architecture- and application-aware power savings recipes for large-scale HPC applications achieving up to 21% and 32% energy savings on HPC production applications run at scale. Additional work focused on the memory sub-system involved our methodology that uses application and machine characterization information to build predictive machine learning models that can accurately quantify phase-level sensitivity to the reduced per core memory bandwidth resulting from changes in the memory bus frequency to reduce the power. We evaluated the predictive capability of the model on real applications and validated them at a fine grain level by looking at 43 individual computational phases or application hotspots as well as the whole application. For more than 91% of the application hotspots, the prediction error is less than 10% (15) . Building from these validated performance and power models collaborations among SUPER team members developed an automated end-to-end system to reduce the complexity of developing and deploying machine learning models for performance, power, and energy. The new framework Automatic Multi-objective Modeling with Machine Learning (AutoMOMML) enabled multi-objective optimizations (power and performance) for HPC workloads.
- Research Organization:
- Univ. of California, San Diego, CA (United States)
- Sponsoring Organization:
- USDOE Office of Science (SC), Advanced Scientific Computing Research (ASCR). Scientific Discovery through Advanced Computing (SciDAC)
- DOE Contract Number:
- SC0006620
- OSTI ID:
- 1481285
- Report Number(s):
- DOE-UCSD-0006620; DE-FG02-11ER26049
- Country of Publication:
- United States
- Language:
- English
Similar Records
Aggregation and Structuring of Materials and Chemicals Data from Diverse Sources
HAM: Hotspot-Aware Manager for Improving Communications with 3D-Stacked Memory