Power-efficient user-adaptable solutions for mobile embedded devices
MetadataShow full item record
Advancements in Computing Systems, with increasing computing power in smaller frames, such as mobile embedded devices, have lead to the support of multitude of applications with increased functionality and greater complexity, thus causing significant battery drain. Limited improvement in battery technology over the years, has forced system designers to utilize the available energy efficiently thus making energy management one of the foremost concerns in mobile devices. Analysis reveals that users differ in context and resource usage patterns which can be leveraged for power savings thus highlighting the need for power-efficient user-adaptable solutions. This dissertation explores architectural and user-level solutions to manage the energy consumed in mobile embedded devices. At the architectural-level, we develop a statistical model for energy estimation in mobile devices using performance counters by correlating power estimates with component access statistics. Evaluation of the model shows a strong correlation between performance parameters and power consumption and the model predicts power consumption for newer observations with minimal errors. Towards power management, predictive models for dynamic cache reconfiguration are developed by analyzing the relationship between cache parameters and power consumption for individual applications. These models are technology-independent and perform better than some brute-force approaches, such as traversing through a list of cache configurations which also incurs significant simulation time and overhead. Further, we extend the energy models to a multi-core environment by developing core-level, shared-memory models for chip multiprocessors. At the user-level, we emphasize the need for user-adaptable solutions by conducting a user study on a smartphone dataset. Using the dataset, statistical battery-aware usage models are developed to understand the impact of individual sub-systems on battery drain. Further, we analyze past idle behavior of users and use a time series approach to predict future idle behavior. This information can be used within the context of power management, to reduce idle power consumption. In addition, with increasing sensitivity of user data, providing context-adaptive security for efficient allocation of resources is a plus. Towards this goal, we formulate the problem and propose a greedy heuristic that maximizes security depending on user context while meeting a given power budget. Further, we develop a novel security-aware energy allocation scheme where security-critical tasks are prioritized over normal tasks depending on the context. The models developed here are cross-validated to demonstrate their accuracy. In addition, power estimates obtained from the model are supplemented with power analysis of varying factors affecting power consumption in single and multi-core systems.