Performance Characterization of a Commercial Video Streaming Service

摘要: Despite the growing popularity of video streaming over Internet, problems such as re-buffering and high startup latency continue to plague users. In this paper, we present an end-to-end characterization Yahoo's service, analyzing 500 million chunks downloaded a two-week period. We gain unique visibility into causes performance degradation by instrumenting both CDN server client player at chunk level, while also collecting frequent snapshots TCP variables from network stack. uncover range issues, including asynchronous disk-read timer cache misses server, variability in network, buffering delays dropped frames client. Looking across same session, or destined IP prefix, see how some are relatively persistent, depending on video's popularity, distance between client's operating system, browser, Flash runtime.