Monitoring the Cray XC30 Power Management Hardware Counters

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract

This paper explores the usefulness of the data provided by the power management (PM) hardware counters available on the Cray XC30 platform. PM data are collected for two molecular modelling codes, DL POLY and CP2K, both of which are run over multiple compute nodes. The first application is built for three programming environments (Cray, Intel and gnu): hence, the data collected is used to test the hypothesis that the choice of compiler should not impact energy use significantly. The second code, CP2K, is run in a mixed OpenMP/MPI mode, allowing us to explore the relationship between energy usage and thread count.

The Cray-compiled DL POLY code had the lowest energy usage on average, 3-4% lower than the Intel and gnu results. In general, energy usage follows execution time. For the CP2K code, an energy-usage sweet spot of three threads per MPI process was revealed; for higher thread counts, execution times increase monotonically with energy usage.
Original languageEnglish
Title of host publicationCUG2015 Proceedings
PublisherCUG
Number of pages7
Publication statusPublished - 1 Apr 2015

Fingerprint

Dive into the research topics of 'Monitoring the Cray XC30 Power Management Hardware Counters'. Together they form a unique fingerprint.

Cite this