The traditional transmission medium for continuous media applications has been a circuit switched network with a fixed channel data rate, carrying an analog signal with no provision for error control, thus delivering a variable quality output depending on instantaneous signal information content and channel quality. The first video coders were designed to output a constant bit rate (CBR) stream to match this fixed bandwidth, disregarding any information redundancy inherent in the encoded media. Digitization of audio and video streams however opens up the possibility of employing compression techniques that minimize bandwidth requirements by exploiting media intrinsic and human perception properties[9], thus producing a constant quality but variable bit rate (VBR)[19] digital stream. This makes packet switching the desired mode of network transport due to the advantages of statistical multiplexing which can realize large economies of scale by supporting constant quality rather than constant bandwidth service.
Traditional continuous media distribution packages all the media components (audio and video) in a single composite wide stream. It is possible to mimic this behavior in a packet switched environment by appending audio data to video frames[20], but this approach prevents us from treating each media component in the most appropriate way during transmission. By transmitting each type of media (and going further, substreams of each media type) independently, not only is media specific treatment allowed, but there is also the potential for diminishing the effects of communication problems. For instance, congestion can be controlled by employing a protocol supporting priorities (based on the type of information transferred on each stream), that drops the least important packets, rather than arbitrary ones, thus maximizing perceived reception quality. Such schemes have been proposed for B-ISDN networks using the Asynchronous Transfer Mode (ATM)[21][22]. Since graceful video quality degradation is deemed to be more acceptable than audio degradation, such a congestion control scheme can first drop video packets; this can be taken further by using hierarchical coding (see Section ii-C) and similarly prioritizing the media substreams. Unfortunately, most packet networks do not explicitly take into account any timing relations among separate transported streams, and thus additional protocol mechanisms have to be provided to achieve intermedia synchronization for playback.
Many techniques commonly used in packet network protocols are based on the assumption that data communication is not particularly delay sensitive but is extremely error sensitive. This assumption does not hold for the real time, high volume, interactive traffic discussed above. For instance, error free transmission, considered essential for most existing data communication applications, is commonly implemented by adding error detection information to transmitted streams and retransmiting any lost or corrupted data. The delay requirements of continuous media applications however may not permit retransmissions, especially for long delay transcontinental or satellite links. Thus, the issue now is the actual error rate for the end-to-end transport mechanisms employed and how it can be dealt with. Since audio and video streams are less sensitive than other data streams to errors, with the perceived reception quality depending heavily on the encoding and compression schemes employed as well as on the channel characteristics, one proposal for error control is to rely on increasing transmission overhead by including error correction rather than error detection information. Whatever the scheme, it is more profitable to formulate the error and delay requirements as soft real time ones, i.e., statistical delay and error rate guarantees such as percentile of packets or bits received correctly within a delay bound[13]. Depending on the exact model that is chosen for defining the application's quality of service requirements, several approaches have been proposed for actually guaranteeing it. Note that quality of service provision is still very much a research area[23].
An issue that has been long debated is the choice between connection oriented and connectionless services. In many packet networks the connectionless approach has been selected for the network layer, with information travelling in datagrams, since this stateless model seems to fit better the bursty nature of the traffic and is also more resilient with respect to network reconfigurations that lead to routing changes. Additional semantics are added, if required, on the transport layer service, by using an appropriate protocol. However, to support the real time service guarantees that continuous media need, it may be required to reserve resources or make other advance arrangements before media transmission starts. The relatively high cost for such set-up coupled with the long life of such communication sessions in terms of packets transferred, seems to fit connection oriented models better. Naturally, in connection oriented networks, such advance preparations can be supported directly by the network layer.
To actually make any service guarantees to continuous media applications, it may be necessary to depart from the best effort service provided in many networks. This service is a direct consequence of the fact that every packet presented to the network boundaries is admitted without regard to the impact it may have on the congestion and quality of service that existing connections are experiencing. To keep the quality of service at least close to the level that these applications have bargained for, admission control mechanisms will have to be introduced that first determine the effects that a new session will have on the existing ones and then admit the new traffic only if the impact is not negative[16][21][22]. Note that routing decisions, resource reservations and admission decisions are interrelated, since the network may be able to support a request by making appropriate reservations over some routes only. Again, these considerations argue in favor of adopting some form of connection establishment scheme, where routing may be fixed in advance, along with resource reservations.
During data transfer, traffic policing mechanisms may have to be used to ensure that applications behave as promised with regard to the data traffic they produce[16]. In addition, traffic shaping mechanisms can help regulate the data transmission so that sensitive data meet their deadlines without overrunning the receivers abilities (flow control) and temporary overload situations in the network are dealt with gracefully (congestion control).