Title:  Time-Shifted Prefetching and Edge-Caching of Video Content: Insights, Algorithms, and Solutions

Committee:

Dr. Raghupathy Sivakumar, ECE, Chair, Advisor

Dr. , Co-Advisor

Dr. Faramark Fekri, ECE

Dr. Douglas Blough, ECE

Dr. Karthik Sundaresan, ECE

Dr. Umakishore Ramachandran, CoC

Abstract: Video traffic accounts for 82% of global Internet traffic and is growing at an unprecedented rate. As a result of this rapid growth and popularity of video content, the network is heavily burdened. To cope with this, service providers have to spend several millions of dollars for infrastructure upgrades; these upgrades are typically triggered when there is a reasonably sustained peak usage that exceeds 80% of capacity. In this context, with network traffic load being significantly higher during peak periods (up to 5 times as much), we explore the problem of prefetching video content during off-peak periods of the network even when such periods are substantially separated from the actual usage-time. To this end, we collected YouTube and Netflix usage from over 1500 users spanning at least a one-year period consisting of approximately 8.5 million videos collectively watched. We use the datasets to analyze and present key insights about user-level usage behavior, and show that our analysis can be used by researchers to tackle a myriad of problems in the general domains of networking and communication. Thereafter, equipped with the datasets and our derived insights, we develop a set of data-driven prediction and prefetching solutions, using machine-learning and deep-learning techniques (specifically supervised classifiers and LSTM networks), which anticipates the video content the user will consume based on their prior watching behavior, and prefetches it during off-peak periods. We find that our developed solutions can reduce nearly 35% of peak-time YouTube traffic and 70% of peak-time Netflix series traffic. We also developed and evaluated a proof-of-concept system for prefetching video traffic.