Category-Based YouTube Request Pattern Characterization

Media content distribution systems make extensive use of computational resources, such as disk and network bandwidth. The use of these resources is proportional to the relative popularity of the objects and their level of replication over time. Therefore, understanding request popularity over time can inform system design decisions. As well, advertisers can target popular objects to maximize their impact. Workload characterization is especially challenging with user-generated content, such as in YouTube, where popularity is hard to predict a priori and content is uploaded at a very fast rate. In this paper, we consider category as a distinguishing feature of a video and perform an extensive analysis of a snapshot of videos uploaded over two 24-hour periods. Our results show significant differences between categories in the first 149 days of the videos’ lifetimes. The lifespan of videos, relative popularity and time to reach peak popularity clearly differentiate between news/sports and music/film. Predicting popularity is a challenging task that requires sophisticated techniques (e.g. time-series clustering). From our analysis, we develop a workload generator that can be used to evaluate caching, distribution and advertising policies. This workload generator matches the empirical data on a number of statistical measurements.