Techniques and tools for measuring energy efficiency of scientific software applications

16th International workshop on Advanced Computing and Analysis Techniques in physics research

物理学;计算机科学

Abdurachmanov, David^1 ; Elmer, Peter^2 ; Eulisse, Giulio^3 ; Knight, Robert^4 ; Niemi, Tapio^5 ; Nurminen, Jukka K^6 ; Nyback, Filip^6 ; Pestana, Gonçalo^5,6 ; Ou, Zhonghong^6 ; Khan, Kashif^5,6

Digital Science and Computing Center, Faculty of Mathematics and Informatics, Vilnius University, Vilnius, Lithuania^1

Department of Physics, Princeton University, Princeton

NJ

08540, United States^2

Fermilab, Batavia

IL

60510, United States^3

Research Computing, Office of Information Technology, Princeton University, Princeton

NJ

08540, United States^4

Helsinki Institute of Physics, PO Box 64, Helsinki

FI-00014, Finland^5

Aalto University, PO Box 11100, Aalto

00076, Finland^6

关键词: Hardware and software;    High performance computing (HPC);    High-throughput computing;    Measurement techniques;    Scientific computing environments;    Scientific softwares;    Software optimization;    Techniques and tools;   

Others  :  https://iopscience.iop.org/article/10.1088/1742-6596/608/1/012032/pdf
DOI  :  10.1088/1742-6596/608/1/012032

学科分类：计算机科学（综合）

来源: IOP

【 摘 要 】

The scale of scientific High Performance Computing (HPC) and High Throughput Computing (HTC) has increased significantly in recent years, and is becoming sensitive to total energy use and cost. Energy-efficiency has thus become an important concern in scientific fields such as High Energy Physics (HEP). There has been a growing interest in utilizing alternate architectures, such as low power ARM processors, to replace traditional Intel x86 architectures. Nevertheless, even though such solutions have been successfully used in mobile applications with low I/O and memory demands, it is unclear if they are suitable and more energy-efficient in the scientific computing environment. Furthermore, there is a lack of tools and experience to derive and compare power consumption between the architectures for various workloads, and eventually to support software optimizations for energy efficiency. To that end, we have performed several physical and software-based measurements of workloads from HEP applications running on ARM and Intel architectures, and compare their power consumption and performance. We leverage several profiling tools (both in hardware and software) to extract different characteristics of the power use. We report the results of these measurements and the experience gained in developing a set of measurement techniques and profiling tools to accurately assess the power consumption for scientific workloads.