Titlebar

Export bibliographic data
Literature by the same author
plus on the publication server
plus at Google Scholar

 

Analysis and Modeling of the Energy Consumption of DVFS Processors for Parallel Scientific Computing

Title data

Stachowski, Matthias:
Analysis and Modeling of the Energy Consumption of DVFS Processors for Parallel Scientific Computing.
2017
Event: ISC High Performance 2017 , 18.06.2017 - 22.06.2017 , Frankfurt.
(Conference item: Conference , Poster )

Abstract in another language

The topic of this dissertation is the analysis of the energy consumption of applications from the area of scientific computing on parallel hardware and the analytical modeling of the resulting power and energy consumption.
To do so, multi-core DVFS processors are considered. The dissertation starts with the investigation of shared memory systems, and will later be extended to cluster systems with distributed address space, which requires to additionally capture the energy consumption of communication.
The ultimate goal is to reduce energy consumption as far as possible without sacrificing the runtime performance.
Early results are based on investigations of the power and energy consumption of the SPEC CPU benchmarks and the PARSEC and SPLASH benchmark suites for different DVFS processors.
Although the energy consumption depends on the execution time, it can be observed that the lowest execution time achieved for a specific number of threads and operational frequency does not necessarily lead to the lowest energy level. This allows us to conclude that there is an application-specific power consumption profile that also requires attention. An analytical model is used as a starting point, distinguishing between the static and the dynamic power consumption resulting from computations.
This model allows us to derive an a priori approximation of the optimal frequency and number of cores to be used for a given processor.
For sequential applications like the SPEC benchmarks, this approach seems to work quite well. We will also investigate the coherence between the internal execution characteristics of applications and their consumed energy.
The analytical model will be extended step by step to also capture memory access costs and communication.
Although some metrics have already been proposed to capture the power and energy efficiency together with the runtime performance, such as the performance per watt or the energy-delay product, we have started to define some new energy metrics. Since both runtime performance and energy consumption are considered together, also metrics combining them are needed.
So our metrics include speedup and reduction factors for the execution time and the energy in isolation as well as a combination of them. These metrics depend on two
variables – the number of cores and the frequency which can be changed on multi-core DVFS processors at runtime.
Especially the relative power increase factor (RPI) is a promising and easy-to-use-metric for the combination of energy consumption and execution time depending on these two variables.
For measuring the energy consumption the Running Average Power Limit (RAPL) Interface of the Intel architecture is used.
These RAPL sensors can be accessed by control registers, known as Model Specific Registers (MSR). An easy to extend framework for automatic measuring and evaluation of applications is in development.
Furthermore, we experiment with multidimensional curve/surface fitting methods, such as regression and least-square methods, to obtain a mathematical model for a given application.
With these techniques we only have to measure some processor settings of an application to be able to recreate the energy consumption data of the other settings. The first results are promising and until now we measure less than half of the energy points to get good results which safes a lot of time.

Further data

Item Type: Conference item (Poster)
Refereed: Yes
Keywords: Modeling; Energy Consumption; DVFS; Parallel Computing
Institutions of the University: Faculties > Faculty of Mathematics, Physics und Computer Science > Department of Computer Science > Chair Applied Computer Science II > Chair Applied Computer Science II - Univ.-Prof. Dr. Thomas Rauber
Result of work at the UBT: Yes
DDC Subjects: 000 Computer Science, information, general works
000 Computer Science, information, general works > 004 Computer science
Date Deposited: 10 Dec 2019 14:30
Last Modified: 10 Dec 2019 14:30
URI: https://eref.uni-bayreuth.de/id/eprint/39099