Analysis and modeling of YouTube traffic

Abstract

YouTube currently accounts for a significant percentage of the Internet’s global traffic. Hence, understanding the characteristics of the YouTube traffic generation pattern can provide a significant advantage in predicting user video quality and in enhancing network design. In this paper we present a characterization of the traffic generated by YouTube when accessed from a regular PC. Based on this characterization, a YouTube server traffic generation model is proposed, which, for example, can be easily implemented in simulation tools. The derived characterization and model are based on experimental evaluations of traffic generated by the application layer of YouTube servers. A YouTube server commences the download with an initial burst and later throttles down the generation rate. If the available bandwidth is reduced (e.g., in the presence of network congestion), the server behaves as if the data excess that cannot be transmitted due to the reduced bandwidth were accumulated at a server’s buffer, which is later drained if the bandwidth availability is recovered. As we will show, the video clip encoding rate plays a relevant role in determining the traffic generation rate, and therefore, a cumulative density function for the most viewed video clips will be presented. The proposed traffic generation model was implemented in a YouTube emulation server, and the generated synthetic traffic traces were compared to downloads from the original YouTube server. The results show that the relative error between downloads from the emulation server and the original server does not exceed 6% for the 90% of the considered videos.