Media streaming, which has emerged as a key application of the Internet, allows users to continually play media data out as it arrives from a distant server. When a user requests a media stream, the server begins transmitting packets containing the desired data to the client. As the data arrives, the client begins to play it out as audio and video sequences. The system is sensitive to the availability time of the successive frames of media data at the client. If the next frame of video or audio is not available at the client by the time it is due to be played out, playout cannot continue.
Because media packets are forwarded across the Internet in a best effort manner, however, they may be dropped, they may contain errors and their time en-route from server to client may vary. To combat these problems, streaming media systems prebuffer data at the client before playout begins. Buffering averages the behavior of the channel over time. It smoothes out delay variations and allows time for retransmission attempts to replace packets that are dropped or that contain errors.
A natural question, then, concerns the length of this preroll buffer. While a long preroll buffer will provide greater immunity to variations in the channel, it will also increase the delay between the time a media stream is requested and the time that playout begins at the client. Adaptive playout schemes have arisen as a method to reduce these delays by allowing shorter buffer lengths while also maintaining a low probability that channel disruptions will halt playout at the client. With these schemes, the rate at which audio and video is played out is varied according to the fullness of the client buffer. The schemes presuppose that variation in playout rate is un-noticeable or subjectively superior to halting playout.
In this paper we analyze average playout rate and probability of underflow performance for adaptive playout schemes. We extend the analysis available in previous works by providing a more generalized channel model. In Yuang [2], the channel is treated as a medium in which the packet interarrival times are randomly distributed. There is no treatment for lost packets or retransmissions. In Steinbach [1], packets arrive at deterministic intervals but may randomly contain errors. In this work we will model random interarrival times and random packet errors. We will consider three cases: short media clips, long programs streamed at the maximum available bandwidth, and long programs streamed at 90% of the maximum available bandwidth, allowing room for retransmissions.
We organize the paper as follows. In section 2 we review what has been done in the literature. In section 3 we describe the extensions to these introduced in this work. Next, in section 4 we will present the results that we obtained for the three scenarios described above using the analysis. Unresolved issues in the analysis will be discussed in section 5.