Data Driven Smartphone Energy Level Prediction

The body of mobile applications is growing at a near-exponential rate; many applications are increasing in both scale, complexity, and their demand for energy. The energy density of smartphone batteries is increasing at a comparably insignificant rate, and thus inhibits the practicality of these applications. Despite the importance of energy to mobile applications, energy is rarely considered by mobile applications developers primarily due to lack of knowledge about users and the absence of energy-aware developer tools. We conduct a large-scale user study to measure the energy consumption characteristics of 15500 BlackBerry smartphone users. Our dataset is several orders of magnitude larger than any previous work. Using this dataset we build the Energy Emulation Toolkit (EET) that allows developers to evaluate the energy consumption requirements of their applications against real user energy traces. The EET computes the successful execution rate of energy-intensive applications across all users, specific devices, and specific smartphone user types. We also consider active adaptation to energy constraints. By classifying smartphone users based on their charging characteristics we demonstrate that energy level can be predicted within 72% accuracy a full day in advance, and through an Energy Management Oracle energy intensive applications can adapt their execution to achieve a near optimal successful execution rate.