Cut the Tail: Mobile Energy Saving Using Radio Tail Prediction

There are growing number of people using mobile computing devices such as Smart phones and tablets, to run different types of network applications. The cellular radio interface for communication in those network applications can cause significant energy drains. This is mainly because of the RaidoTails when the cellular radio remains in high energy state after communication has ended. To solve this problem, many existing works have discussed to employ Fast Dormancy, which is a feature that can force client radio to quickly go into a low energy state. The key problem in this line of research is to predict the proper time to invoke Fast Dormancy. A proper time means that when we invoke Fast Dormancy we would save as much idle time in high energy state as possible for radio cellular and also be able to avoid high overhead caused by frequent state transitions. There are some existing works studying this problem and use data mining algorithms such as decision tree to predict the right time of End of Session(EOS) event in packet transmission. However, they only focus on applications running in background and cannot learn a general model for different types of network applications with all possible user actions. Thus our aim is to design a simple and general algorithm that can predict the proper time for Fast Dormancy in any applications with all possible user actions. We propose our effective and efficient algorithm to cut the RadioTail in this paper. We first analyze the characteristics of existing work. Then we propose a framework to reduce the dimensionality of the features used in prediction. After that, we apply the reduced features in different prediction algorithms. Lastly, we evaluate our proposed method on a data set tracked from Android OS in 18 days. Results show that our proposed algorithm can achieve 84% accuracy, which is more than 10% higher than the baseline.