基於最佳EDP之動態電壓頻率調整技術與實現

Design and Implementation of an EDP-based Dynamic Voltage and Frequency Scaling Mechanism

陳則翔(Tse-Hsiang Chen 陳則翔)

指導教授 ： 梁文耀

國立臺北科技大學/電資學院/資訊工程系研究所/碩士(2012年)

https://doi.org/10.6841/NTUT.2012.00637

摘要

在手持式裝置的廣泛運用下，能源消耗的控管將是一個相當重要的議題，文中所採用的是眾多省電技術的其中一種「DVFS(動態頻率調節機制)」，在運行中調節至適當的頻率來達到省電的效果；往往的觀念中，只要將其頻率調至最低可達到最低能量消耗，卻忽略了調降頻率會拉長總執行時間，反而會花費這一段額外的能量消耗，而所省下的能量又小於這段額外花費的能量，才會有此一現象， 而文中採用的EDP指標正是用來權衡能量消耗與執行時間的關係；故本文中提出一個MAR-CSEDP近似方程式，透過此方程式能將我們當前的MAR值找到適當頻率，獲得最佳的EDP值[1]，達到能量消耗和效能損失的降低，而這一個頻率點我們稱之為CSEDP。
在本篇論文中使用我們提出的MAR-CSEDP近似方程式，實作出EPS Governor運行在XScal PXA270上，利用MiBench這套測試程式，評估我們的演算法，最後以EDP值做為評估優劣依據，經結果顯示，我們的EPS Governor 根據MAR-CSEDP方程式較優於Ondemand Governor[2]，其EDP值可下降幅度從0.429%~14.596%。

關鍵字

低功耗 ； 動態電壓頻率調整 ； 記憶體存取率 ； 中央處理器使用率 ； EDP

並列摘要

Energy saving is an important issue, especially for the handheld devices. The dynamic voltage and frequency scaling (DVFS) is one of the mechanisms in the energy saving research area. DVFS can be used to reduce energy consumption of a processor while tasks are running. However, reducing the frequency for a processor might not induce minimized energy consumption, since the execution time will be extended due to the frequency being decreased and the program will consume extra energy during this extended period. As a result, the induced extra energy could be higher than what has been saved.
Experiment results indicated that the execution time is an important factor for the energy consumption. The EDP (energy and delay product) metric[1] has been used to evaluate the trade-off between energy consumption and execution time. In this thesis, an approximation equation based on the correlation of the memory access rate and the critical speed for the minimum EDP metric is conducted for frequency and voltage prediction. The correlation equation can thus be deduced from the relationship and used at task execution time to find a frequency. According to the correlation equation, an appropriate frequency (called CSEDP) can be found to minimize the EDP.
The energy saving DVFS algorithm based on the correlation equation is presented in this thesis. The algorithm has been implemented on Linux as a kernel-space power manager. The experiment result shows that our algorithm performed better than the commonly used Linux Ondemand Governor[2]. It indicates that the EDP values are lower than that of Ondemand from 0.429% to 14.596% with 5.27% in average.

並列關鍵字

Low Power ； Energy saving ； DVFS ； CPU utilization ； EDP metric

參考文獻

[1] M. Horowitz, T. Indermaur and R. Gonzalez, "Low-power digital design," IEEE Symposium, Low Power Electronics, 10-12 Oct. 1994, pp.8-11.

[2] V. Pallipadi and A. Starikovskiy, "The ondemand Governor," Linux Symposium, Shared-Subtree Concept, Implementation, and Applications in Linux, 2006, pp.215-230.

[3] F. Kluge, S. Uhrig, J. Mische, B. Satzger and T. Ungerer, "Dynamic Workload Prediction for Soft Real-Time Applications," IEEE 10th International Conference on Computer and Information Technology (CIT), 2010, pp.1841-1848.

[4] L. Hollevoet, A. Dewilde, K. Denolf, F. Catthoor and F. Louagie, "A power optimized display memory organization for handheld user terminals," Design, Automation and Test in Europe Conference and Exhibition, vol. 3, 2004, pp. 294-299.

[5] V. Spiliopoulos, S. Kaxiras and G. Keramidas, "Green Governors: A framework for continuously adaptive DVFS," International Green Computing Conference and Workshops (IGCC), 2011, pp. 1-8.