Internet video comprises a major portion of Internet traffic today. While core
and access network capacities continue to grow, optimizing Internet video delivery
will remain a challenge, as new forms of video and technology keep emerging, and
content publishers continue to seek higher Quality of Experience(QoE) of users due
to its correlations with user engagement and revenue. The goals of this thesis are to
create a deeper understanding of the Internet video ecosystem and to propose novel
methodologies to improve QoE of Internet video delivery.
In this thesis, we make the following contributions. First, we create a deeper understanding of video management plane by characterizing it, at scale, along its key
dimensions based on more than 100 content publishers data spanning 27 months, and
we propose new metrics to measure complexity of video management plane. Next,
in order to enhance video control plane, we propose Oboe, a system that improves
the dynamic range of Adaptive Bitrate(ABR) algorithms by automatically tuning
ABR behaviors to the current network state of a client connection to improve QoE
of a wide range of users. Through testbed experiments, we show Oboe significantly
improves performance of several different ABR algorithms. Finally, given that performance of ABRs critically depends on throughput prediction accuracy, we propose
a new throughput prediction approach, called Xatu, to address challenges in existing
prediction methods used by ABRs. Xatu, a learning based throughput prediction
framework, uses richer information (e.g., ISP or chunk size) without apriori partitioning data, and we show that Xatu reduces the prediction error by more than 23%
relative to state-of-the-art throughput prediction.