A Mathematical Method for Online Autotuning of Power and Energy Consumption with Corrected Temperature Effects

详细信息    查看全文

• 作者：Reiji Sudaa ^{reiji@is.s.u-tokyo.ac.jp} ; Luo Chenga ; Takahiro Katagirib
• 关键词：Autotuning ; Power ; Energy ; Temperature ; Cost Modeling ; Experimental Design
• 刊名：Procedia Computer Science
• 出版年：2013
• 出版时间：2013
• 年：2013
• 卷：18
• 期：Complete
• 页码：1302-1311
• 全文大小：197 K

文摘

This paper discusses a method of online autotuning of power or energy consumption. A problem of online power modeling is the strong positive influence of the temperature on the power consumption. Without proper treatment of the temperature, an autotuning mechanism will misunderstand a variant measured at a low temperature to be power-efficient. This paper proposes a model of the relation between power and temperature, a Bayesian inference formula to estimate the power consumption with estimates of estimation errors, and an experimental design (i.e. algorithm of choosing variants) for online autotuning. The proposed methods are evaluated by simulation, and applied to online power and energy optimizations.