Learning Battery Consumption of Mobile Devices

We introduce a data-driven model to predict battery consumption of apps. The state-of-the-art models used to blame battery consumption on apps are based on microbenchmark experiments. These experiments are done under controlled conditions where power measurements of each internal resource (CPU, Bluetooth, WiFi, ...) are readily available. We empirically verify that such models do not capture the power consumption behavior of mobile devices in the wild and propose instead to train a regression model using data collected from logs. We show that this learning approach is correct in the sense that under mild assumptions, we can recover the true battery discharge rate of each component. Finally, we present experimental results where we consistently outperform a model trained on micro-benchmarks.