Further advances in networking coupled with rapid advances in storage technologies will make it feasible to build multimedia on-demand servers capable of providing services similar to those of a neighborhood videotape rental store on a metropolitan-area network. In this research we investigate architectures for multimedia servers that can permit multiple subscribers to interactively record and playback multimedia objects over integrated computer networks. We address two main requirements in the development of such a multi-subscriber multimedia server: (1) Continuity of playback of each media constituting a multimedia object, and (2) mutual synchrony of playback of the different media constituting a multimedia object. In order to address the continuity requirement, we develop admission control algorithms with which a multimedia server can satisfy the maximum number of subscribers simultaneously without violating continuity of playback for any of them. One such algorithm is a quality proportional scheme that retrieves media units at a rate proportional on an average to their respective playback rates, but uses a staggered toggling technique by which successive numbers of media blocks retrieved are fine tuned individually to admit and service an optimal number of subscribers simultaneously. Initial performance estimation shows that the quality proportional policy is an order of magnitude more scalable compared to straightforward admission control policies such as servicing one subscriber per disk head and round robin servicing of subscribers. During retrieval of a multimedia object, mediaphones that playback the different media components of the object may have recording and playback rate mismatches. Furthermore, emerging integrated networks such as ATM are expected to introduce jitter that can nevertheless be bounded in the transport of media units from servers to mediaphones. In order to ensure synchronous playback of multimedia objects, we study a feedback technique in which a multimedia server uses lightweight messages called feedback units generated by mediaphones and transmitted back to the server to detect asynchronies during playback. We study various resynchronization policies such as aggressive, conservative, and probabilistic and compare their performance for video/audio playback. Initial performance estimation shows that whereas aggressive policies are expected to perform best only at lower playback rates, conservative policies are expected to perform best only at higher playback rates. In contrast, probabilistic policies have the potential to perform uniformly well at all playback rates. The policies and algorithms developed in this project form the basis for a prototype multimedia server being developed at the UCSD Multimedia Laboratory.
