Analysis on the state of mobile HTTP video streaming at the client-side

With the popularity of the HTTP video streaming in mobile networks, the network operators are increasingly interested in the performance of the video streaming at the client-side. However, unlike content providers, network operators are not privy to the client-side or server-side, thus they have no good access to estimating the performance from the video state at the client-side. To address this limitation, this paper proposes a scheme to investigate the video state from the network operators´ perspective. Our study is based on the network packets collected from live DC-HSPA+ network of China Unicom at the client-side, and makes mainly two contributions. First, a Deep Packet Parsing (DPP) method is proposed to decode real video packets into frames. On basis of parsed frame information, the user-viewing behaviors combined with video re-buffering events are recognized with accuracy close to 100%. Second, a Bayesian Classifier is trained based on the video labels generated from DPP process to identify the re-buffering events, and the accuracy is close to 80%. Our proposed methods can be directly used by network operators since the input parameters of the Bayesian Classifier are only statistics ones of video packets which can be easily collected in real time at network side.