適用於嵌入式系統輸出入事件之耗能剖析工具

Abstract

近年來，耗能的問題日漸受到重視，低功耗已成為設計嵌入式系統的一大重點，而耗能剖析工具則是一套協助使用者評估耗能的工具，雖然存在不少針對處理器的耗能剖析工具，但隨著雲端計算的趨勢，手持裝置主要的耗電來源不再是處理器而是周邊的輸出入裝置，例如：無線網路裝置、液晶顯示器。因此，本論文從軟體的角度，因應雲端計算的趨勢，提出一套輸出入裝置之耗能剖析架構可適用於量測和耗能模型的評估方法，並實現於具作業系統之嵌入式平台。此工具克服輸出入裝置非同步與分散收集的行為，把輸出入事件與對應的軟體相互關聯，依據此關係將硬體上的耗能歸回軟體中的程序層級和函數層級，讓軟體開發者得知其軟體在輸出入裝置的耗能行為，進而做更進一步的調校。實驗結果顯示，在無線網卡和液晶顯示器透過程序層級的耗能模型所評估的耗能與實際量測之耗能相比，其誤差率都小於4%。

Recently, energy issues become more and more important. Low power has become one of critical design issues in embedded systems. The energy profiling tools are to assist users estimate energy consumption. So far, many energy profilers for processors have been proposed. However, with the trend of cloud computing, I/O devices become the most power-consuming devices such as WLAN and LCD. Hence, this thesis proposes a generic energy profiling framework for I/O devices from software perspective. The framework can be applied to measurement-based and power model-based approaches to estimate energy consumption. The proposed tool was implemented on the XScale PXA270 embedded platform with the Linux kernel 2.6.15. This tool overcomes asynchronous I/O issue and scatter-gather issue to correlate energy consumption of I/O events with process level and function level. This tool enables software designers to understand energy impact of their software, and software designers can make further optimization. The experimental results show the error rate of total estimated energy calculated by the proposed process-level energy models for WLAN and TFT-LCD is within 4%.