Energy consumption of CUDA kernels with varying thread topology

Special Issue Paper
  • 211 Downloads

Abstract

The energy consumption and energy awareness of modern GPGPU devices becomes important with large GPGPU based system installations. Measurements of the average power consumption have been done and their predictions are reported in literature. However, by observing several repeatable impacts on energy consumption within our experiments we conclude that the available models are limited to ideal scheduling behavior. This conclusion results from relating the noticed impacts to the scheduling mechanisms on GPGPUs. Past work assumed that the consumed energy is considered to be linearly dependent on the thread count, but as we show this is only valid if perfect scheduling is feasible. We demonstrate this by revealing nonlinear increases of energy consumption in several particular cases. Thus we conclude that linear models for predicting the energy consumption are not always reliable.

Keywords

CUDA Energy awareness Energy consumption Energy efficiency GPGPU 

Notes

Acknowledgements

We would like to thank Matthias Noack and Florian Wende for valuable discussions. This work is funded by the German Bundesministerium für Bildung und Forschung (BMBF) project ENHANCE, grant No. 01IH11004A-G.

References

  1. 1.
    Collange S, Defour D, Tisserand A (2009) Power consumption of GPUs from a software perspective. In: Proceedings of the 9th international conference on computational science, Part I (ICCS ’09). Springer, Berlin, pp 914–923 Google Scholar
  2. 2.
    Haifeng W, Qingkui C (2012) An energy consumption model for GPU computing at instruction level. Int J Adv Comput Technol 4(2):192–200 Google Scholar
  3. 3.
    Hong S, Kim H (2010) An integrated GPU power and performance model. Comput Archit News 38(3):280–289 CrossRefMathSciNetGoogle Scholar
  4. 4.
    Huang S, Xiao S, Feng W (2009) On the energy efficiency of graphics processing units for scientific computing. In: Proceedings of the 2009 IEEE international symposium on parallel & distributed processing (IPDPS ’09), IEEE Computer Society, Washington, DC, USA, pp 1–8 CrossRefGoogle Scholar
  5. 5.
    Kozin I (2010) Energy efficiency investigation (power & performance). In: PRACE workshop “New languages & future technology prototypes. http://www.prace-ri.eu/PRACE-Workshop-New-Languages Google Scholar
  6. 6.
    McIntosh-Smith S, Wilson T, Ibarra AA, Crisp J, Sessions RB (2012) Benchmarking energy efficiency, power costs and carbon emissions on heterogeneous systems. Comput J 55(2):192–205 CrossRefGoogle Scholar
  7. 7.
    Corp NVIDIA (2009) NVIDIA’s next generation CUDA compute architecture: Fermi. Whitepaper Google Scholar
  8. 8.
    NVIDIA Corp (2012) NVML API reference manual Google Scholar
  9. 9.
    Rofouei M, Stathopoulos T, Ryffel S, Kaiser W, Sarrafzadeh M (2008) Energy-aware high performance computing with graphic processing units. In: Proceedings of the 2008 conference on power aware computing and systems (HotPower’08), USENIX Association, Berkeley, pp 11–19 Google Scholar
  10. 10.
    Top500Org (2011) TOP500 list—November 2011. http://www.top500.org/lists/2011/11

Copyright information

© Springer-Verlag 2012

Authors and Affiliations

  1. 1.Zuse Institute BerlinBerlinGermany

Personalised recommendations