For Peer ReviewTwo Decades of Internet Video Streaming:A Retrospective View•1:13generated by users. The sheer number of UGC objects is orders of magnitude higher than that oftraditional movies or TV programs, and evolves rapidly. As of March 2013 [Statistics 2013], everysecond, 1.2 hours worth of video is uploaded on YouTube by users around the world, attracting almost140 views for every person in the world on average. Earlier industry insiders estimated that YouTubespent over$1million a day to pay for its server bandwidth. This has defeated any effort towardsincreasing server capacity and improving the user experience. Saxenaet al.have revealed that theaverage service delay of YouTube is nearly6.5seconds, which is much longer than the other measuredsites [Saxena et al. 2008]. While peer-to-peer mechanisms would be a candidate to scale the video streaming system, the hugenumber of videos with a highly skewed popularity implies that many of the peer-to-peer overlays。 willbe too small to function well. Moreover, user generated videos are generally short (70% are shorterthan 1 minute, even though YouTube has relaxed the length limit [Cheng et al. 2012]), implying thatan overlay will suffer from。 an extremely high churn rate. These factors together make existing per-video based overlay design suboptimal, if not entirely inapplicable. Fig. 7. A sample graph of YouTube videos and their links. On the other hand, the UGC nature introduces new social relations and interactions among videosand users. In particular, there are interesting relations among the YouTube videos: the links to relatedvideos generated by uploader's choices form a small-world network, as illustrated in Fig. 7 [Chenget al. 2012]. This suggests that the videos have strong correlations with each other. A user oftenquickly loads another related video when finishing the previous one. An earlier work that exploressuch relationships is NetTube [Cheng and Liu 2009], which introduces an upper-layer overlay on topof the swarms of individual videos. In the upper-layer overlay, given a peer, neighborhood relations areestablished among all the swarms that contain this peer. This conceptual relation facilitates a peer toquickly locate the potential suppliers for the next video and enable a smooth transition. The relationsalso enable effective pre-fetching of videos. While the repository of short videos is large, the next videoin YouTube is most likely confined by the related list of the current video. This list in general includes20videos at most。
正在翻譯中..
