A Video Streaming Technique over TCP

A video streaming application needs to employ a transport layer protocol to transmit packetized media data. Since TCP is the most dominant protocol over the Internet, it is reasonable to consider TCP for video streaming. However, the AIMD behavior and retransmission timeouts of TCP may introduce unacceptable end-to-end delay, and therefore it is difficult to provide good quality of streaming video. While it is widely believed that TCP is not suitable for video streaming, it has been reported that video streaming over TCP is a feasible solution when some amount of buffering delay is acceptable. This is because the receiver-side buffering can accommodate the rate variability of TCP. This research is devoted to investigate the feasibility of TCP for video streaming when receiver buffering is employed. Our starting point is an analytic model of a video streaming system over TCP which is characterized by the steady state throughput analysis. Based on this model, we derive a feasibility condition to achieve a desired performance. Specifically, we consider the performance in terms of the buffer underrun probability, since buffer underruns generally cause playback disruptions and degrade perceptual video quality. Analytic characterization of the feasibility condition states that the required buffer size is determined by packet loss rate, round-trip time (RTT), and a buffer underrun probability. Since the Internet does not provide any information about network conditions in advance, we also develop a realtime algorithm to estimate packet loss rate, RTT, and the buffering delay. We evaluate the performance by streaming an encoded video generated by an error resilient scalable codec. Experimental results show that the performance of a video streaming system without receiver buffering exhibits significant rate variability and quality fluctuation. However, a system coupled with the proposed algorithm is capable of reducing the rate variability of TCP and improving overall video quality significantly, when buffering delay given by the feasibility condition is acceptable.