Optimizing the video transcoding workflow in content delivery networks

The current approach to transcoding in adaptive bit rate streaming is to transcode all videos in all possible bit rates which wastes transcoding resources and storage space, since a large fraction of the transcoded video segments are never watched by users. To reduce transcoding work, we propose several online transcoding policies that transcode video segments in a "just-in-time" fashion such that a segment is transcoded only to those bit rates that are actually requested by the user. However, a reduction in the transcoding work should not come at the expense of a significant reduction in the quality of experience of the users. To establish the feasibility of online transcoding, we first show that the bit rate of the next video segment requested by a user can be predicted ahead of time with an accuracy of 99.7% using a Markov prediction model. This allows our online algorithms to complete transcoding the required segment ahead of when it is needed by the user, thus reducing the possibility of freezes in the video playback. To derive our results, we collect and analyze a large amount of request traces from one of the world's largest video CDNs consisting of over 200 thousand unique users watching 5 million videos over a period of three days. The main conclusion of our work is that online transcoding schemes can reduce transcoding resources by over 95% without a major impact on the users' quality of experience.