Adaptive performance sensitivity model to support GPU power management

Francesco Paterna, Ujjwal Gupta, Raid Ayoub, Umit Ogras, Michael Kishinevsky

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

Abstract

Integrated graphics units consume a large portion of power in client and mobile systems. Pro-active power management algorithms have been devised to meet expected user experience while reducing energy consumption. These techniques often rely on power and performance sensitivity models that are constructed at design phase using a number of workloads. Despite this, the lack of representative workloads and model identification overhead adversely impact accuracy and development time, respectively. Conversely, two main challenges limit runtime post-design identification: the absence of sensitivity feedback from the system and the limited computational resources. We propose a two-stage methodology that first identifies the features of the sensitivity model offline by leveraging a reduced amount of training data and then uses recursive least square algorithm to fit and adapt the coefficients of the model to workload changes at runtime. The proposed adaptive approach can reduce offline training data by 50% with respect to full offline model identification while maintaining accuracy as much as 95% on average.

Original languageEnglish (US)
Title of host publication1st Workshop on AutotuniNg and aDaptivity AppRoaches for Energy efficient HPC Systems, ANDARE 2017 - A Workshop part of PACT 2017
PublisherAssociation for Computing Machinery
VolumePart F132205
ISBN (Electronic)9781450353632
DOIs
StatePublished - Sep 9 2017
Event1st Workshop on AutotuniNg and aDaptivity AppRoaches for Energy efficient HPC Systems, ANDARE 2017 - Portland, United States
Duration: Sep 9 2017 → …

Other

Other1st Workshop on AutotuniNg and aDaptivity AppRoaches for Energy efficient HPC Systems, ANDARE 2017
CountryUnited States
CityPortland
Period9/9/17 → …

Fingerprint

Identification (control systems)
Energy utilization
Graphics processing unit
Power management
Feedback

Keywords

  • Adaptive learning
  • Frame time
  • GPU
  • Online modeling
  • Performance sensitivity
  • Power management

ASJC Scopus subject areas

  • Human-Computer Interaction
  • Computer Networks and Communications
  • Computer Vision and Pattern Recognition
  • Software

Cite this

Paterna, F., Gupta, U., Ayoub, R., Ogras, U., & Kishinevsky, M. (2017). Adaptive performance sensitivity model to support GPU power management. In 1st Workshop on AutotuniNg and aDaptivity AppRoaches for Energy efficient HPC Systems, ANDARE 2017 - A Workshop part of PACT 2017 (Vol. Part F132205). [a5] Association for Computing Machinery. https://doi.org/10.1145/3152821.3152822

Adaptive performance sensitivity model to support GPU power management. / Paterna, Francesco; Gupta, Ujjwal; Ayoub, Raid; Ogras, Umit; Kishinevsky, Michael.

1st Workshop on AutotuniNg and aDaptivity AppRoaches for Energy efficient HPC Systems, ANDARE 2017 - A Workshop part of PACT 2017. Vol. Part F132205 Association for Computing Machinery, 2017. a5.

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

Paterna, F, Gupta, U, Ayoub, R, Ogras, U & Kishinevsky, M 2017, Adaptive performance sensitivity model to support GPU power management. in 1st Workshop on AutotuniNg and aDaptivity AppRoaches for Energy efficient HPC Systems, ANDARE 2017 - A Workshop part of PACT 2017. vol. Part F132205, a5, Association for Computing Machinery, 1st Workshop on AutotuniNg and aDaptivity AppRoaches for Energy efficient HPC Systems, ANDARE 2017, Portland, United States, 9/9/17. https://doi.org/10.1145/3152821.3152822
Paterna F, Gupta U, Ayoub R, Ogras U, Kishinevsky M. Adaptive performance sensitivity model to support GPU power management. In 1st Workshop on AutotuniNg and aDaptivity AppRoaches for Energy efficient HPC Systems, ANDARE 2017 - A Workshop part of PACT 2017. Vol. Part F132205. Association for Computing Machinery. 2017. a5 https://doi.org/10.1145/3152821.3152822
Paterna, Francesco ; Gupta, Ujjwal ; Ayoub, Raid ; Ogras, Umit ; Kishinevsky, Michael. / Adaptive performance sensitivity model to support GPU power management. 1st Workshop on AutotuniNg and aDaptivity AppRoaches for Energy efficient HPC Systems, ANDARE 2017 - A Workshop part of PACT 2017. Vol. Part F132205 Association for Computing Machinery, 2017.
@inproceedings{ff435561465c432881f68d55072f7a26,
title = "Adaptive performance sensitivity model to support GPU power management",
abstract = "Integrated graphics units consume a large portion of power in client and mobile systems. Pro-active power management algorithms have been devised to meet expected user experience while reducing energy consumption. These techniques often rely on power and performance sensitivity models that are constructed at design phase using a number of workloads. Despite this, the lack of representative workloads and model identification overhead adversely impact accuracy and development time, respectively. Conversely, two main challenges limit runtime post-design identification: the absence of sensitivity feedback from the system and the limited computational resources. We propose a two-stage methodology that first identifies the features of the sensitivity model offline by leveraging a reduced amount of training data and then uses recursive least square algorithm to fit and adapt the coefficients of the model to workload changes at runtime. The proposed adaptive approach can reduce offline training data by 50{\%} with respect to full offline model identification while maintaining accuracy as much as 95{\%} on average.",
keywords = "Adaptive learning, Frame time, GPU, Online modeling, Performance sensitivity, Power management",
author = "Francesco Paterna and Ujjwal Gupta and Raid Ayoub and Umit Ogras and Michael Kishinevsky",
year = "2017",
month = "9",
day = "9",
doi = "10.1145/3152821.3152822",
language = "English (US)",
volume = "Part F132205",
booktitle = "1st Workshop on AutotuniNg and aDaptivity AppRoaches for Energy efficient HPC Systems, ANDARE 2017 - A Workshop part of PACT 2017",
publisher = "Association for Computing Machinery",

}

TY - GEN

T1 - Adaptive performance sensitivity model to support GPU power management

AU - Paterna, Francesco

AU - Gupta, Ujjwal

AU - Ayoub, Raid

AU - Ogras, Umit

AU - Kishinevsky, Michael

PY - 2017/9/9

Y1 - 2017/9/9

N2 - Integrated graphics units consume a large portion of power in client and mobile systems. Pro-active power management algorithms have been devised to meet expected user experience while reducing energy consumption. These techniques often rely on power and performance sensitivity models that are constructed at design phase using a number of workloads. Despite this, the lack of representative workloads and model identification overhead adversely impact accuracy and development time, respectively. Conversely, two main challenges limit runtime post-design identification: the absence of sensitivity feedback from the system and the limited computational resources. We propose a two-stage methodology that first identifies the features of the sensitivity model offline by leveraging a reduced amount of training data and then uses recursive least square algorithm to fit and adapt the coefficients of the model to workload changes at runtime. The proposed adaptive approach can reduce offline training data by 50% with respect to full offline model identification while maintaining accuracy as much as 95% on average.

AB - Integrated graphics units consume a large portion of power in client and mobile systems. Pro-active power management algorithms have been devised to meet expected user experience while reducing energy consumption. These techniques often rely on power and performance sensitivity models that are constructed at design phase using a number of workloads. Despite this, the lack of representative workloads and model identification overhead adversely impact accuracy and development time, respectively. Conversely, two main challenges limit runtime post-design identification: the absence of sensitivity feedback from the system and the limited computational resources. We propose a two-stage methodology that first identifies the features of the sensitivity model offline by leveraging a reduced amount of training data and then uses recursive least square algorithm to fit and adapt the coefficients of the model to workload changes at runtime. The proposed adaptive approach can reduce offline training data by 50% with respect to full offline model identification while maintaining accuracy as much as 95% on average.

KW - Adaptive learning

KW - Frame time

KW - GPU

KW - Online modeling

KW - Performance sensitivity

KW - Power management

UR - http://www.scopus.com/inward/record.url?scp=85038400540&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85038400540&partnerID=8YFLogxK

U2 - 10.1145/3152821.3152822

DO - 10.1145/3152821.3152822

M3 - Conference contribution

AN - SCOPUS:85038400540

VL - Part F132205

BT - 1st Workshop on AutotuniNg and aDaptivity AppRoaches for Energy efficient HPC Systems, ANDARE 2017 - A Workshop part of PACT 2017

PB - Association for Computing Machinery

ER -