Web-based video distribution services are powerful tools that enterprises use to attract and retain customers, and to advertise merchandise over the Internet. Consequently, Web-based video distribution service providers can benefit from understanding the characteristics and quality of shared video workloads. How can service providers implement efficient storage techniques and improve access to voluminous worldwide video content generated by users?

This paper investigates the workload characteristics of video sharing on the Web in an effort to recommend effective system design for providers of Web-based video distribution services. The authors selected four providers of Web-based video distribution services and analyzed traces of metadata from nearly two million videos with over six billion views. They examine the cumulative distributions of ratings, comments, and uploads of videos by users. Video popularity is ascertained from the cumulative distributions of the total viewers and their ratings. Zipf’s law [1] and power laws [2] are used to illuminate the hyperbolic behaviors of the video-sharing services, while Pareto distributions [3] are used to examine their invariant workload characteristics.

The authors used datasets from Dailymotion, Metacafe, Yahoo, and Veoh to characterize Web-based video sharing workloads. A comparison of the video workloads of these four services reveals that most users don’t rate videos unless they are popular; comment on or bookmark favorite videos; upload videos, except for when advertising house goods, music, or news, or promoting healthy habits; or upload longer videos, unless they are privileged users.

The authors observe that perceptions of the popularity of videos viewed and rated by users exhibit Pareto distributions. Consequently, they used these factors to define lifetime measures of the popularity of individual videos. Twenty percent of the most popular videos in the study accounted for at least 85 percent of the total videos watched and at least 80 percent of the video ratings. Unfortunately, the authors could not establish a generic parametric power law model to explain the distributions of video popularity across the Web-based video distribution services.

However, the authors identify two important invariant characteristics of Web-based video distribution services: the infrequent ratings of, and comments on, videos with social interaction tools; and the smaller number of users who upload videos compared to users who watch videos. They creatively use graphs to illuminate the relationships between computer memory caching efficiencies and the popularity of video content, without deriving any prediction equations. Prediction equations would be useful for allocating memory resources and scheduling frequently accessed videos during peak periods.

The research results in this paper should encourage interesting debates among Web-based video distribution services and academics. Information storage and retrieval experts and operating systems researchers should participate in this discussion.