基於高階層硬體原件使用率以離線和半在線方式的智慧型手機耗電剖析

Abstract

如何對智慧型手機子系統作精確的耗電剖析是必要且廣泛的研究議題，當前智慧型手機子系統的耗電剖析是基於透過分析在不同電源狀態運行單一硬體原件子系統的功耗構造而得到耗電模型。精確的功率曲線可以導致高效的電源模式，然而，現有的電源分析技術(如在線耗電剖析) 因為沒有考慮到智慧型手機的耗電行為及一些如電池劣化或充放電狀態等關鍵性問題所造成的影響，仍然會導致顯著的功率估計錯誤。此外，現有的電力分析技術是無法適用於所有類型的智慧型手機的，例如，使用外接電源監視器來測量功耗的離線電力剖析無法量測配備不可拆卸電池的智慧型手機。因此，對於智慧型手機耗電剖析是有進步空間的。 本文首先提出了名為符號回歸與分群(symbolic regression and clustering, SRC)的離線耗電剖析技術, 透過分析智慧型手機的非線性、非同步及非同構三種功耗行為改善耗電剖析的準確性。與其它傳統基於剖析使用率的耗電剖析技術相比，SRC可提供更好且更準確的耗電資訊。此外，透過Eureqa工具的輔助，SRC可以產生正確表示硬體原件子系統耗電行為的耗電模型。接著本文提出了一個對於多核處理器能夠剖析更多細的技術以剖析特定的子系統及半在線剖析技術用以剖析裝有不可拆卸電池的智慧型手機。透過分析充電資料及電池劣化、非同步行為、充電狀態影響三個關鍵問題，半在線技術改善了傳統在線耗電剖析技術。與傳統傳統在線耗電剖析技術相比，半在線耗電剖析技術可減少約84％的錯誤率。最後，本文提出了EPRO ，不僅能正確地作耗電剖析，還指出了導致智慧型手機低能量使用率的耗電錯誤所在，使得應用程式開發者可以知道在哪裡，以及如何解決耗能問題。

Accurate power consumption estimates of commercial smartphone subsystems, power profiles, is necessary for wide research areas. Typically, power profiles is represented as the relation between the operating state of individual hardware subsystems and power consumption. To accurately obtain the power consumption estimates, power profiling, is the key factor to have an efficient power profiles. However, existing power profiling techniques, such as online power profiling, still cause significant power estimate errors because they do not address complete power consumption behaviors or some critical issues such as battery capacity degradation. Moreover, the existing power profiling techniques are not practical enough to apply for all type of commercial smartphones, e.g., the offline-based power profiling, using external power monitors to measure the power consumption, cannot be applied for all types of smartphones, equipped with non-removable batteries, or are not flexible enough to apply for generating power profiles of a large number of smartphones, such as cloud-based testing. Hence, there is a gap to improve the power profiling techniques which can be applied for various commercial smartphones. To have better power profiling techniques, this dissertation first proposes the offline power profiling technique, named as SRC shorted for symbolic regression and clustering, which improves the accuracy of utilization-based power profiling approach by addressing the power consumption behaviors in smartphones, nonlinear and asynchronous behaviors, which affect the accuracy of power estimates. By addressing those power behaviors, SRC can capture more details of power profiles, compared with the other traditional utilization-based profiling approach. Second, this dissertation proposes a semi-online power profiling technique, namely SEMI, which aims for estimating power consumption of smartphones for which the external power meters cannot be applied, e.g., the non-removable batteries. By using charging data and addressing three critical issues, i.e., battery capacity degradation, asynchronous behaviors, and state-of-charge effect, SEMI can improve the traditional online power estimation techniques. In particular, SEMI can reduce about 84% errors compared with the traditional online power estimations. Finally, to show the application of the generated power profiles, this dissertation proposes energy productivity-based scheduling (EPS), a novel scheduling method for schedule the background sync jobs when smartphones are in the active state. EPS aims to increase the energy productivity while maintaining the total energy usage in smartphones. EPS shows the significant improvement of energy productivity of 3G network while reducing its energy consumption on the background sync jobs generated by heavily using foreground apps.