Evaluation of the HPC Applications Dynamic Behavior in Terms of Energy Consumption

Vyhodnocení dynamického chování HPC aplikací z pohledu energetické spotřeby

conference paper

English

Vysocký Ondřej, Ing.
Beseda Martin (FIT)
Říha Lubomír, doc. Ing., Ph.D. (FIT)
Zapletal Jan
Nikl Vojtěch, Ing.
Lysaght Michael (FIT)
Kannan Venkatesh (FIT)

READEX, energy efficient computing,MERIC, RADAR, HDEEM, RAPL, runtime tuning, Haswell processor

This paper introduces the READEX project tuning approach which exploits the dynamic application behavior and its potential for energy savings. The paper is focused on themanual applications evaluation from the energy consumption optimisation point of view. As an examples we have selected one complex application, the ESPRESO library and two simplified applications from the ProxyApps benchmark tool suite. ESPRESO containsmany types of operations including I/O, communication, sparse BLAS and dense BLAS. The results show that static savings are 5.6-12.3% and dynamic savings are 4.7-9.1%. The highest total savings for ESPRESO are 21.4% as a combination of 12.3% static savings and 9.1% dynamic savings. The ProxyApp applications, Kripke and Lulesh, were presented for two configurations each. The first configuration of the Kripke saved 29.3% energy, almost only by static tuning. On the other hand, the second configuration shows us only 18.8% savings, but over a third of it was saved by dynamic switching CPU core and uncore frequencies. The Lulesh test cases saved 28.9%, respectively 26.7%.

This paper introduces the READEX project tuning approach which exploits the dy-
namic application behavior and its potential for energy savings. The paper is focused
on the manual applications evaluation from the energy consumption optimisation point
of view. As an examples we have selected one complex application, the ESPRESO
library, and two simplified applications from the ProxyApps benchmark tool suite.
ESPRESO contains many types of operations including I/O, communication, sparse
BLAS and dense BLAS. The results show that static savings are 5.6-12.3 % and dy-
namic savings 4.7-9.1 %. The highest total savings for ESPRESO are 21.4 % as a com-
bination of 12.3 % static savings and 9.1 % dynamic savings.
The ProxyApp applications Kripke and Lulesh, were presented for two configu-
rations each. The first configuration of the Kripke saved 29.3 % energy, almost only
by static tuning. On the other hand, the second configuration shows us only 18.8 %
savings, but over a third of it was saved by dynamic switching CPU core and uncore
frequencies. The Lulesh test cases saved 28.9 %, respectively 26.7 %.

2017

30–49

PROCEEDINGS OF THE FIFTH INTERNATIONAL CONFERENCE ON PARALLEL, DISTRIBUTED, GRID AND CLOUD COMPUTING FOR ENGINEERING

Civil-Comp Proceedings

978-1-905088-66-9

Civil-Comp Press

Stirlingshire

10.4203/ccp.111.3

2-s2.0-85020398791