Large scale clustered continuous media (CM) servers deployed in applications like video-on-demand have high availability requirements.In the event of server failure, streams from the failed servers must be reassigned to healthy servers with minimum service disruption.Such servers may also suffer from periods of transient overload resulting from a high degree of customer interactivity.For example, in a video-on-demand system if a large number of users are viewing a favorite game, many of them can simultaneously request a replay of an interesting part of the game.This requires a large number of '5nteractive" channels within a short period of time and can result in a transient server overload.In this paper we propose solutions for graceful recovery from overload scenarios arising out of server failure or customer interactions.Rapid resource reclamation is key to overload tolerance, and our proposed solution is based on rate adaptive stream merging and content insertion techniques, We also utilize conventional time-sharing techniques to handle transient overload.We show that while merging is necessary for achieving overload tolerance, it is not sufficient, and for a complete solution, content insertion is required.Specifically, we consider a general clustered CM server architecture model where multiple servers can fail simultaneously.We develop a model for resource shortfalls that occur as a result of overload on failure.We also describe optimal polynomial time algorithms for recovering resources to the maximum extent possible, by clustering streams in real time.

Concurrent retrieval of continuous media from a physical storage device can be achieved by interleaving data and providing a suitable scheduling algorithm. Scheduling approaches that exploit gains from statistical multiplexing are susceptible to a nonzero probability of frame loss due to the variable-bit-rate characteristic of compressed video. With interframe encoding schemes (such as specified by the MPEG standard), the losses propagate, resulting in a net loss of frames that exceeds the fraction of lost data. In this paper, we describe a mechanism for the storage and retrieval of MPEG-encoded video from a single disk storage system. The scheme balances the need for the reliable delivery of MPEG frames with the desire to support the largest number of sessions. Our approach reorganizes the MPEG-encoded video stream based on the relative importance of the frames and maps them to the storage device geometry. The reorganization reduces the impact of frames lost due to missed deadlines and distributes the frame losses over time and among sessions. Simulation results show that the new approach improves performance when compared to conventional storage and scheduling schemes.

Video-on-demand represents a key demonstrative application for enabling multimedia technology in communication, database, and interface research.This application requires solving a number of diverse technical problems including the data synchronization problem for time-dependent data delivery.In this paper we describe the general requirements of video-ondemand and introduce a system supporting content-based retrieval and playback for the structure and content of digital motion pictures.In our model we capture domain-specific information for motion pictures and provide access to individual scenes of movies through queries on a temporal database.We describe our implementation of this service using existing workstation and storage technology.