Research On Network Resource Allocation And Transmission Scheduling For New Streaming Video Services

Crowdsourced live video broadcasting(livecast)and 360-degree video have be-come increasingly popular in recent years.Comparing with conventional video services,these new streaming services have the characteristics of rich interactions,super-high res-olution,and massive transcoding.They face severe challenges in efficient resource al-location and high-quality network transmission.However,Previous studies on livecast platforms lack a thorough analysis of danmu interactions between viewers and broad-casters.They allocate capacities based on channels' popularity,and can not accom-modate the heterogeneous Quality-of-Experience(QoE)demands for massive users.Current works on streaming 360 videos have not done enough on revealing the rela-tionship between the chunk prefetching and the viewport prediction errors,which leads to inefficient bandwidth utilization.This dissertation focus on solving the problems of low transmission efficiency and unreasonable resource allocation in new streaming video services.Specifically,it includes research on modeling viewer interactive behav-ior,optimizing the video transcoding and delivery strategies in livecast scenarios and the 360° video chunks transmission scheduling in virtual reality(VR)scenarios.The contributions in this work are summarized as follows:We study the problem of characterizing and modeling novel interactions in new streaming video services.In livecast and 360 VR scenarios,viewers are allowed to perform rich interaction such as posting comments,donating virtual gifts and changing the viewpoint when watching the videos.Understanding user behavioral patterns is essential for improving Internet video streaming services.To this end,we carry out a systematic study on characterizing and modeling viewer interactions.On one hand,for livecast services,we apply empirical models to capture three important aspects of the interactions,namely the gift donating process,the comment-posting and gift-sending activities,and the channel's popularity.Rather than simply relying on curve fitting,we also interpret the model parameters with real-world meanings.We further analyze the influences of the broadcaster's behavioral factors on a channel's popularity,and present a random forests-based methodology for popularity predicting.On the other hand,for 360°interactive videos,we apply multiple methods to predict the user viewing direction and the further bandwidth.Different from existing studies that focus on improving prediction accuracy,we build a probabilistic model for describing the temporal and spatial distributions of both viewpoint and bandwidth prediction errors.We study the problem of optimizing service capacity allocation among crowd-sourced livecast channels.Previous livecast studies use viewer population to mea-sure a channel's importance,and prioritize popular channels when allocating service capacities.However,such popularity-based service strategies are not applicable in the business model that relies heavily on monetary donations,and viewer's heterogenous QoE demands cannot be satisfied.To solve this problem,we present a danmu inter-action aware livecast framework.We propose an algorithm that balances channels'popularity,profitability,and cloud instances'rental cost when determining the video representation sets and the rented instance locations for transcoding.We also develop a model to assess viewer perceived service quality,propose a practical strategy to sched-ule viewer requests between collaborative transcoding-enabled edge nodes,and provide each viewer with proper streaming bitrate and delivery path.Experiments driven by real-world measurement data show that compared with current strategy,our proposed solution can effectively improve the system utility by 88.1%and the average viewer QoE by 76.2%.We study the problem of fault-tolerant prefetch scheduling in adaptive 360-degree video streaming.The adaptive streaming needs to select the quality of every video tiles according to the estimated user viewpoint and network capacity.The prefetch decisions are affected by both the viewpoint prediction error and bandwidth prediction error.Un-like existing solutions that only consider the unilateral impact,we propose a robust 360-degree video transmission mechanism when neither viewport and bandwidth can be accurately predicted.We leverage the constructed probabilistic model of viewport to estimate the values of 360 video chunks,and further prefetch chunks in a fault-tolerant way.We present a value-driven optimization framework that captures the tradeoffs between the goal of improving the picture quality(along with the balance of quality distribution)in potential field of views and the goal of retrieving chunks ahead of play-back deadlines so as to maximize the expected viewer QoE.We propose a heuristic and a deep reinforcement learning based approach to make prefetch decisions.Extensive trace-driven experiments demonstrated our framework can enhance viewer experience,and outperforms the state-of-the-art methods by at least 40.7% on average.