In order to enable exascale computing, concepts for substantial energy savings are required. Dynamic voltage and frequency scaling (DVFS) is widely known to provide suitable energy saving potentials. However, the customarily utilized DVFS mechanism of the Linux kernel determines clock frequencies solely based on an idle time analysis. In contrast to this, we use an empirical approach based on preparatory measurements of the energy consumption at all available frequencies. From the resulting data we deduce energy-optimal frequencies, which are used in subsequent production runs. The described methodology can be deployed with routine granularity to account for varying code characteristics. For evaluation purposes, the approach is applied to the UG4 numerical simulation software. First results exhibit an average energy saving potential of approximately 10% while increasing the runtime by about 19%.
%0 Journal Article
%1 noKey
%A Dick, Björn
%A Vogel, Andreas
%A Khabi, Dmitry
%A Rupp, Martin
%A Küster, Uwe
%A Wittum, Gabriel
%D 2015
%I Springer Berlin Heidelberg
%J Computing and Visualization in Science
%K DVFS Energy HLRS empirical
%N 2
%P 89-97
%R 10.1007/s00791-015-0251-1
%T Utilization of empirically determined energy-optimal CPU-frequencies in a numerical simulation code
%U http://dx.doi.org/10.1007/s00791-015-0251-1
%V 17
%X In order to enable exascale computing, concepts for substantial energy savings are required. Dynamic voltage and frequency scaling (DVFS) is widely known to provide suitable energy saving potentials. However, the customarily utilized DVFS mechanism of the Linux kernel determines clock frequencies solely based on an idle time analysis. In contrast to this, we use an empirical approach based on preparatory measurements of the energy consumption at all available frequencies. From the resulting data we deduce energy-optimal frequencies, which are used in subsequent production runs. The described methodology can be deployed with routine granularity to account for varying code characteristics. For evaluation purposes, the approach is applied to the UG4 numerical simulation software. First results exhibit an average energy saving potential of approximately 10% while increasing the runtime by about 19%.
@article{noKey,
abstract = {In order to enable exascale computing, concepts for substantial energy savings are required. Dynamic voltage and frequency scaling (DVFS) is widely known to provide suitable energy saving potentials. However, the customarily utilized DVFS mechanism of the Linux kernel determines clock frequencies solely based on an idle time analysis. In contrast to this, we use an empirical approach based on preparatory measurements of the energy consumption at all available frequencies. From the resulting data we deduce energy-optimal frequencies, which are used in subsequent production runs. The described methodology can be deployed with routine granularity to account for varying code characteristics. For evaluation purposes, the approach is applied to the UG4 numerical simulation software. First results exhibit an average energy saving potential of approximately 10% while increasing the runtime by about 19%.},
added-at = {2015-11-12T16:09:31.000+0100},
author = {Dick, Björn and Vogel, Andreas and Khabi, Dmitry and Rupp, Martin and Küster, Uwe and Wittum, Gabriel},
biburl = {https://puma.ub.uni-stuttgart.de/bibtex/2eb68df7362f6017c3fd049c2cb4f9a29/bjoern_dick},
doi = {10.1007/s00791-015-0251-1},
interhash = {ee872f8891616773765cf6cfd757389f},
intrahash = {eb68df7362f6017c3fd049c2cb4f9a29},
issn = {1432-9360},
journal = {Computing and Visualization in Science},
keywords = {DVFS Energy HLRS empirical},
language = {English},
number = 2,
pages = {89-97},
publisher = {Springer Berlin Heidelberg},
timestamp = {2016-03-07T16:42:29.000+0100},
title = {Utilization of empirically determined energy-optimal CPU-frequencies in a numerical simulation code},
url = {http://dx.doi.org/10.1007/s00791-015-0251-1},
volume = 17,
year = 2015
}
