Abstract
In this thesis, we focus on large time-scale scheduling of mobile data transfer, e.g., in minutes or hours. Such large time-scale provides significantly more room for performance improvement in real-life scenarios, which differs our work from most existing channel-aware scheduling studies. In particular, we design, implement, and evaluate a framework for profiling and scheduling based on Markov decision theory, using the Android platform. Our trace-driven simulations show that mobile users in real-life scenarios can benefit significantly from our framework. For example, 50% of mobile users will enjoy 20%-90% throughput improvement with a deadline guarantee of 40 minutes and an average delay of 15 minutes. In addition, we quantify and reduce the overhead of generating model parameters of the proposed scheduling algorithms. We determine the best training window size: 30 days. We adopt various clustering algorithms to group users in order to reduce training overhead. We empirically determine the best system parameters of the clustering algorithms using real traces. Our clustering algorithms reduce the training overhead without sacrificing too much performance: it saves up to 59.9% of training time while incurring <18% performance degradation.