Boosting Power Efficiency of HPC Applications with GEOPM
Event Type
Computer Architecture
Education and Training
Energy consumption
Performance Analysis and Optimization
Performance Tools
TimeSunday, June 24th9am - 1pm
DescriptionPower and energy are critical constraints for exascale supercomputing. Optimizing for application performance under such constraints is becoming challenging due to the dynamic phase behavior of scientific codes, increasing variation in processor power efficiency resulting from manufacturing, and due to complexities arising from upcoming heterogeneous architectures and task-based programming models. In order to address some of these challenges, Intel introduced GEOPM, an open-source, hierarchical job-level runtime system to optimize for time-to-solution under a job-level power constraint by leveraging techniques from learning and control systems. GEOPM leverages Intel’s hardware power limiting capability called RAPL. In this tutorial, we will present a high-level overview of the GEOPM architecture, a walkthrough of the GEOPM plugin infrastructure, and demonstrate basic use cases covered by selected tests and example benchmarks packaged with GEOPM. We will also provide a hands-on demonstration of our ECP Power-Steering work incorporated as a third-party plugin into GEOPM. In this part, we will cover algorithms to speed up application critical path through adaptive configuration selection and mitigating process variation through intelligent power management. Additionally, we also discuss best practices for using RAPL and DVFS on production HPC systems.
Content Level Session will cover an introduction, discussion, and interactive demo of GEOPM: a cross-community, extensible, open-source, (BSD 3-clause) power management framework.
Target Audience The tutorial will target general HPC system users, application programmers, and power researchers. It aims to provide a theoretical and practical demonstration of how GEOPM can be used for safe execution and performance optimization of applications running in a power-constrained environment.
PrerequisitesThere are no specific prerequisites that need to be met by the audience. It is preferred (but not necessary) that they carry their laptops so that they can test some of the tutorial codes, benchmarks, and applications discussed during the session.
Postdoctoral Research Staff member
Professor / Chair for Computer Architecture and Parallel Systems at TUM
Principal Engineer