Enhancing Upper-level Performance from Below: Performance Measurement and Optimization in LTE Networks

The current 4G LTE cellular networks provide significantly higher data rate and lower network latency than earlier generation systems. The advances provided by LTE, however, may not be fully utilized by upper-level protocols and applications. In this dissertation, we first investigate the predictability of link bandwidth in cellular networks. We then design and implement a new transport protocol that utilizes the link bandwidth prediction framework. Last, we design and implement a control-theoretic approach for adaptive video streaming for cellular networks. For cellular link bandwidth prediction, we conduct an extensive measurement study in major US commercial LTE networks and identify various lower-level pieces of information that are correlated with cellular link bandwidth. Furthermore, we develop a machine learning based prediction framework named LinkForecast, which uses the identified lower-level information to predict cellular link bandwidth in real time. Our evaluation shows that the prediction is highly accurate: even at the time granularity of one second, the average relative error is 4.7\% under static scenarios and 9.3\% under highly mobile (highway driving) scenarios. The accurate link bandwidth prediction can in turn be utilized by upper-level protocols and applications to improve their performance. We design a new transport protocol LTP that directly utilizes the real-time cellular link bandwidth prediction. Evaluations using real-world traces show that our protocol outperforms traditional TCP protocols, achieving both high throughput and low delay at the same time. Our study highlights the benefits of using radio layer information when designing higher layer protocols for cellular networks. Last, we design and implement a PID control based scheme for adaptive video streaming in cellular networks. We start by examining existing state-of-the-art video streaming adaptation strategies from a control-theoretic point of view, and identifying their limitations. Motivated by the wide adoption and success of PID control in feedback control systems, we develop a PID based adaptation algorithm that optimizes user\u27s perceived QoE in an online manner. Evaluations using large sets of real-world cellular traces show that our solution outperforms existing schemes under various network conditions and different user\u27s QoE preferences. While our approach does not require a very accurate link bandwidth prediction, its performance is further improved when using the accurate link bandwidth predictions from LinkForecast