Mobile User Clustering in Large Time-Scale Data Transfer Scheduling

Ting-An Lin,Yichuan Wang,Cheng-Hsin Hsu,Xin Liu

DOI: https://doi.org/10.1145/2462456.2465714

2013-01-01

Abstract:Mobile data traffic has grown explosively in the past few years, largely due to the increasingly more smartphone users. A large portion of the mobile traffic, such as email and user-generated multimedia contents may tolerate delay to some extent. While there are existing works [1] proposing opportunistic data transfer in small time-scale, they do not leverage the much larger optimization room that is only achievable via large time-scale scheduling. In our previous work, we proposed two large time-scale scheduling algorithms: OSS and OSSL using Markov decision theory [2]. While OSS and OSSL lead to huge improvement on user-selected optimization criteria, such as throughput and energy consumption, they share two limitations, they: (i) need long profiles (e.g., 30 days) for accurate predictions, and (ii) may incur high training complexity (e.g., it takes OSS on average 457.2 sec and 3.96 GB memory to train the model for a single user). In the current work, we eliminate these two limitations via clustering mobile users by the optimization criteria, which is referred to as contexts throughout this abstract.