Energy Modeling and Measurement in High Performance Computing Systems

Date

Oct 16, 2017

Time

10:00 AM - 11:00 AM

Speaker

Thomas Ilsche

Affiliation

ZIH

Language

en

Main Topic

Informatik

Other Topics

Informatik

Description

Energy efficiency is a major criteria for computing in general and High Performance Computing (HPC) in particular. As with any optimization, it is essential to effectively determine the underlying metric: energy consumption. The thesis will discuss the two approaches: modeling the energy consumption and measuring it. However, both energy modeling and measurement commonly require a well-understood accuracy to be useful. To that end, this thesis provides a rigorous evaluation of measurement and modeling techniques. High quality energy measurements are not commonly available and need to be improved to facilitate a broader range of use-cases for energy-efficiency analysis. I demonstrate improvements for the temporal resolution of energy measurements and describe a scalable solution deployed in production HPC systems. Carefully combining these sophisticated energy measurements and application performance traces, gives a detailed insight into application energy-efficiency at full scale HPC systems and down to millisecond- range application regions. The sophisticated measurements also reveal weaknesses in commonly used energy models. I will discuss the fundamental limitations of energy models based on hardware performance counters and suggest application specific energy models as a more accurate alternative. Finally, I propose a concept for a hierarchical energy model that combines different kinds of component models to enable a holistic simulation of future system designs. Together, reliable energy models and measurements are the foundation for the design of energy-efficient systems and their optimal operation. Diese Veranstaltung wird unterstützt von ZIH.

Last modified: Oct 16, 2017, 9:40:13 AM