Abstract
Modern mobile devices integrate numerous power-intensive subsystems and run a wide variety of applications. Traditional power management systems often lack awareness of both the user's desired battery lifetime and the specific application context, resulting in suboptimal performance or premature battery exhaustion. This paper presents BLAST (Battery Lifetime-constrained Adaptation with Selected Target), an Android-based power management framework. By leveraging application-dependent power states (AP-states), BLAST dynamically optimizes performance to meet a user-defined battery lifetime constraint. Implemented entirely in userspace, BLAST requires no OS modifications and demonstrates that battery life can be extended by 25% with negligible impact on user experience.
Introduction: The Gap in Mobile Power Management
Standard mobile power management is fundamentally "blind" in two ways. First, OS-level governors (such as ondemand or powersave) typically react to hardware utilization metrics without context regarding the active application. Second, these systems lack a temporal "deadline" or target lifetime.
For example, a user on a fixed-duration commute might wish to play a game while ensuring their device remains powered until they reach their destination. A standard "powersave" mode might throttle the game to an unplayable state despite sufficient energy, while a "performance" mode might deplete the battery long before the trip ends. BLAST addresses this by treating a target battery lifetime as a primary constraint, maximizing the quality of experience (QoE) within that specific energy budget.
The BLAST Framework
The BLAST framework introduces several key mechanisms to achieve lifetime-constrained adaptation:
Application-Dependent Power States (AP-states)
Power consumption varies significantly between applications even at identical hardware frequencies. BLAST profiles individual apps (e.g., Chrome vs. Angry Birds) to create a lookup table mapping CPU/GPU frequency pairs to average power consumption. This context-aware approach allows for far more precise energy budgeting than generic utilization models.
Heuristic Control and the Energy Tank
BLAST establishes a linear "discharging profile" based on the user's target lifetime and the current battery level. It employs an "energy tank" metaphor: when an application consumes less than its allocated budget, the surplus energy is stored in the "tank" for later use by more demanding tasks. Every second, the Power Controller solves a management problem to identify the highest frequency limits that satisfy the remaining energy budget.
Userspace Implementation and Portability
Unlike many research frameworks that require custom Linux kernels, BLAST is implemented in the Android userspace. It monitors the foreground application and adjusts frequency limits via the sysfs interface. This design ensures high portability across diverse Android hardware without necessitating OS modifications.
Experimental Evaluation and Results
The framework was evaluated using a combination of real-world applications (Google Chrome, Angry Birds), standard benchmarks (Antutu, Geekbench 3, GFX T-Rex), and a user study involving 15 participants on a Qualcomm APQ8064 tablet.
Guaranteed Lifetime and the "25% Sweet Spot"
In testing, BLAST consistently met target battery lifetimes where standard governors failed. A significant finding was the "25% sweet spot": battery life can be extended by 25% with almost no perceptible drop in user satisfaction. For instance, in Temple Run, a 25% extension resulted in a frame rate drop from ~60 to ~50 FPS, which participants rated nearly identical to the unconstrained 0% extension scenario (7.29/10 vs 7.32/10).
Governor Compatibility
BLAST functions as a supervisory layer over existing kernel governors like interactive or ondemand. By dynamically setting upper frequency bounds, it constrains the underlying governor to operate within the calculated energy envelope, ensuring compatibility with existing system optimizations.
Conclusion and Future Work
BLAST demonstrates that application-aware, lifetime-constrained power management is both feasible and effective on modern mobile devices. While the current implementation focuses on CPU and GPU scaling, future work will explore integrating display brightness control—a significant power sink—into the framework. Additionally, refining the profiling process for new applications and accounting for background process energy consumption remain key areas for further development.
Summary updated