An integrated tool for measuring and optimizing performance and energy efficiency

Datum

24.03.2016

Zeit

09:00 - 10:00

Sprecher

Dipl.-Inf. Robert Schöne

Zugehörigkeit

ZIH

Sprache

en

Hauptthema

Informatik

Andere Themen

Informatik

Beschreibung

In recent years, energy efficiency concerns have risen in the field of high performance computing (HPC). While computing performance is still the most important metric for evaluating an HPC system, the total cost of ownership is also considered with energy costs representing a significant part of it. This shift of focus is, for example, reflected by rankings of HPC systems in the order of computations per energy unit as done in the Green500 list and recent SPEC benchmarks. However, even though energy-efficiency metrics are now an accepted part of an HPC system evaluation, they have not yet been an integral part of scalable performance measurement tools. In my thesis, I detail how to include metrics like power, idle state usage, and temperature in application traces, which effectively combines performance and energy efficiency measurement. The recorded information can be used to show whether the system is used efficiently, e.g., whether blocking routines use idle-states and avoid busy-waiting. Furthermore, I describe how the existing measurement infrastructure can be used to influence system characteristics to increase the energy-efficiency of parallel programs. To do so, I describe how power saving features like idle states, performance states and clock modulation can be used and how they influence the performance and power dissipation of systems. Based on this information and the available instrumentation from performance measurement tools, I describe a way to utilize these features and consequently show that performance measurement and energy-efficiency tuning should be combined in a single infrastructure. In my talk, I will give an overview on the state-of-the-art of energy-efficiency measurement in HPC. I further describe the current status of my work, which includes details about energy-efficiency metrics, power-saving features, and energy-efficiency optimizations. Diese Veranstaltung wird unterstützt von Professur für Rechnerarchitektur.

Letztmalig verändert: 24.03.2016, 08:37:14