La loi de convexité énergie-fréquence de la consommation des programmes : modélisation, thermosensibilité et applications

Résumé : Both theoretical and experimental evidence is presented in this work for the existence of an Energy/Frequency Convexity Rule, which relates energy consumption and microprocessor frequency for nanometer-scale microprocessors. Typical nanometer-scale application processors were monitored running specific compute-intensive kernels using high-resolution power gauges. Data gathered during several week-long acquisition campaigns suggest that energy consumed is strongly correlated with the microprocessor’s frequency, and, more interestingly, the curve exhibits a clear minimum over the processor’s frequency range. An analytical model for this behavior is provide and motivated, which fits well with the data. The circumstances are discussed under which this convexity rule can be exploited, and when other methods are more effective, with the aim of improving the microprocessor’s energy efficiency. The Energy/Frequency Convexity Rule is potentially more exploitable by low-power systems, such as battery-powered and embedded systems, and less likely by high-performance computer systems. The Energy/Frequency Convexity Rule is also applied to multi-buddy systems, Amdahl’s law and heterogeneous computing. Given that the microprocessor’s energy consumption is temperature-dependent, a macro-level temperature/power relationship for application processors is introduced and experimentally validated. By adopting a holistic view, this model is able to take into account many of the physical effects that occur within such systems. Via measurements on two pertinent platforms sporting nanometer-scale application processors, it is shown that the power/temperature relationship is indeed very likely exponential over a 20C to 85C temperature range. The data suggest that, for a temperature range between 20C and 55C, a quadratic model is still accurate and a linear approximation is acceptable. Power transformation models are also presented that aim at canceling the temperature biases in power traces. These transformation models are developed to increase the accuracy and meaningfulness of power measurement traces. Besides static power measurements, the transient power and thermal behavior are also analyzed by means of the cooling laws and the temperature/power relationship models. Exponential cooling models are justified for actively-cooled microprocessors. For passively cooled processors however, as frequently found in embedded systems, an exponential law may not be theoretically justified. Here, the tractability of the exact cooling law for a passively-cooled body is analyzed, subject to radiative cooling and a modest level of heat loss via convection. Focusing then on embedded microprocessors, the performance difference between the new passive cooling law and the conventionally-used exponential one is compared. It is shown that, for large surface sizes, the radiative cooling component can be comparable to the convective cooling one. However, for large cooling surface areas of the order of 10 cm2 or more, it is shown that the differences between the passive cooling law and the exponential cooling law are significant. The results thus suggest that, in the absence of accurate temperature measurements, an exponential cooling law is only accurate enough for small-sized SoC systems that require low processing overhead.
Type de document :
Thèse
Informatique [cs]. Telecom ParisTech, 2015. Français
Liste complète des métadonnées

https://hal-mines-paristech.archives-ouvertes.fr/tel-01258577
Contributeur : Claire Medrala <>
Soumis le : mardi 19 janvier 2016 - 11:56:14
Dernière modification le : mardi 12 septembre 2017 - 11:41:25

Identifiants

  • HAL Id : tel-01258577, version 1

Citation

Karel De Vogeleer. La loi de convexité énergie-fréquence de la consommation des programmes : modélisation, thermosensibilité et applications. Informatique [cs]. Telecom ParisTech, 2015. Français. 〈tel-01258577〉

Partager

Métriques

Consultations de
la notice

4374

Téléchargements du document

214