Instead of constant bit rate encoding and delivery with the known problems of variable quality, not fully using the network at all times, and session rejection, we describe a system in which independent video servers, by using a set of fixed quality encodings and analyzing the statistics of the encoded video they are about to deliver, vary the aggressiveness of TCP rate control so as to obtain unequal shares of the contended network, in such a way that more complex video sequences get a bigger share. Compared to an equal share of bit rate, there is much less variation in quality between video streams delivered at the same time. This paper reports on the use of MulTCP, a server side modification to TCP, to apportion bandwidth more appropriate to the demands of the content being streamed. We present results of NS-2 simulations of our autonomous rate adaptive streaming server using MulTCP as the transport and show how similar qualities across content types are maintained when streaming over a contended network.

Video streaming has frequently been deployed using constant bit rate video encoding and transmission as these are easily implemented and network provisioning, although requiring admission control, is straightforward. However, it is sub-optimal for both the user, who experiences time varying quality, and the network operator, who can not fully utilize the network during periods of low usage and is faced with admission control and session refusal at busy times. Instead we propose an equitable quality scheme in which the available network bandwidth is divided between the concurrent sessions so that the same quality is delivered in each. We show that we can increase average quality (mean opinion score) by 0.74, or alternatively increase the number of sessions that can be supported at the same average overall quality by 80%.

Home broadband bandwidth will be a variable commodity as customers run multiple applications through their home hubs simultaneously. The reaction of some of these applications (e.g. Web browsing or e-mail) to congestion or packet loss will be governed by the control mechanisms within the transfer control protocol (TCP/IP). Retransmissions and rate control will slow down TCP applications when in a congested state, but data will, in general, be reliably transmitted. However, applications which are more sensitive to delay, such as voice over IP and videotelephony, are generally sent for delay considerations using the real time protocol (RTP) and the underlying unreliable user datagram protocol (UDP/IP). These protocols deliver timely data but with the downside that any congestion will lead to packet losses which will propagate up to, and need to be managed by, the application layer. This paper reports on efforts to define an application layer quality-of-service framework for the video component of videotelephony over RTP/UDP leading to an initial reference design for specifying videotelephony products. The error resilience features of the latest video-coding standard, H.264, are examined, as are modifications to existing RTP Internet recommendations to allow effective feedback-based repair for this new codec. Results using RTP early feedback and H.264 reference picture selection in a multi-reference configuration will also be presented.

Current generation, large-scale, Internet protocol television (IPTV) systems borrow heavily from the broadcast industry, which makes a number of delivery assumptions that do not apply to IP networks. Consequently we can perceive major improvements if we better match the delivery of IPTV services with the underlying network transport. We can expect next generation IPTV systems to adapt video streams dynamically to maximise throughput, allow constant quality delivery, and degrade gracefully in congested networks. This paper outlines the challenges in optimising IPTV delivery and the contribution that BT's research has made to overcoming some of these.