Power Monitor of Embedded System with Power Signal Compression

嵌入式系統之功率監視系統與功率訊號壓縮

楊政儒(Jheng-Ru Yang)

指導教授 ： 劉靖家

國立清華大學/電機資訊學院/電機工程學系/碩士(2012年)

摘要

即時的功率消耗對於嵌入式系統是重要的資訊，我們可以利用這些資訊最佳
化系統以延長系統使用時間和提高效能。我們使用耗電量測(Power Measurement)
為基礎，直接量測系統運作時的實際功率消耗，也讓軟體得知系統在運行時的實
際的功率消耗，以做耗電最佳化設計。因此，我們採用內嵌式系統即時量測平台
[1]做為功率量測平台並做一些改善。然而，軟體時間和功率量測系統時間是不
同步的，所以當驅動程式發一個命令給功率量測系統，而量測系統接收到這個命
令時間是不一樣的，所以軟體收集的功率消耗資訊是不正確的。因此，我們提出
一個同步化模型去補償驅動程式到功率量測系統的延遲時間。我們的實驗顯示這
段延遲是常數。
除此之外，並不是所有的功率消耗資訊都是有用的，我們只需要觀察功率訊
號的特性，像是某個時間點有比較大的功率消耗或是平均功率，所以我們應用壓
縮感測的壓縮方法來壓縮功率資料。然而，壓縮感測只適用於訊號具有稀疏特性
的訊號，但是並非所有功率訊號都是稀疏訊號。此外，壓縮感測在軟體端必須做
離散餘弦轉換來重建訊號，造成系統整體效能降低。因此，我們提出基於樣板的
壓縮方法來取代壓縮感測的不足。實驗結果顯示，大部份例子採用基於樣板壓縮
方法的壓縮比都比壓縮感測的壓縮比好。最後結合綜合上述兩種方法來壓縮功率
訊號。我們量測處理器並執行 6 種不同程式。實驗結果顯示，給定相對誤差在
15%以下，有 6 個例子壓縮比是小於 0.1，給定相對誤差在 10%以下，有 5 個例
子壓縮比是小於 0.2。而量測記憶體並執行 5 種不同程式。實驗結果顯示，給定
相對誤差在 15%以下，有 5 個例子壓縮比是小於 0.3，給定相對誤差在 10%以下，
有 5 個例子壓縮比是小於 0.4。最後，我們比較壓縮感測和基於樣板壓縮的軟體
效能。實驗結果顯示，樣板壓縮的軟體效能比壓縮感測軟體效能要好。

關鍵字

嵌入式系統 ； 功率量測 ； 壓縮

並列摘要

Real-time Power Consumption in an embedded system is an important reference information
that to optimize endurance and real-time performance. We use measurement-based power profiler
to directly measure the actual power consumption, and also allows software developers to know
the actual power consumption when their programs are executed. So that more effective power
optimization can be done. So I adopt the real-time power the measurements of an embedded
system [1] as our power measurement platform and make some improvements. First, we add
API to let software can start/stop the function to measure power. However, the software time and
power measurement system time are not synchronized, so that the time when power measurement
measure power is not exactly the time when software issues the start command. So we proposed a
synchronous model in which latency between driver to power measurement system is compensated.
Our experiments show that the latency is constant between driver to power measurement system.
Furthermore, because not all power consumption information is meaningful for us and only
certain characteristic are useful. We apply the compressive sensing (CS) compression to compress
power data. However, CS work well to sparse signals. But not all power signals are sparse.
Besides, CS have to do IDCT which result in software overhead. Therefore, we proposed template-
based compression to replace CS. The experimental results show that in most cases template-based
compression ratio (CR) better than CS compression ratio. Finally, we combine CS with template-
based compression to compress power signals and use power measurement for cpu executing 6
benchmarks. The experimental results show that in 6 cases CR is not above 0.1 when acceptable
relative error rate is 15%. In 5 cases CR is not above 0.2 acceptable relative error rate is 10%.
We use power measurement for memory executing 5 benchmarks. The experimental results show
that in 5 cases CR is not above 0.3 when acceptable relative error rate is 15%. In 5 cases CR is
not above 0.4 acceptable relative error rate is 10%. Finally, we compared CS with template-based
1software overhead. The experimental results show that template-based software overhead less than
CS.

並列關鍵字

HASH(0x9dc9afc) ； HASH(0x9dc9b74) ； HASH(0x9dc9bec)

參考文獻

tional Tsing-Hua University, Electrical Engineering Department, 2011.

[2] D. Brooks, V. Tiwari, and M. Martonosi, “Wattch: a framework for architectural-level power

of Computer Science and Engineering, 2009.

[4] D.Donoho, “Compressed sensing,” in IEEE Trans. Information Theory, vol. 52, no. 4, 2006,

pp. 1289–1306.