We focus our work on the storage and retrieval of continuous media data. We have developed a server model supporting the playback operation of continuous media. The architecture of the server model is based on a share-nothing model. All the nodes in the model are independent but connected; they communicate with each other by passing messages through the interconnection network. The server model can be implemented on a distributed memory parallel machine or a collection of personal computers or workstations interconnected by high-speed links.
Our storage strategy employs striping technique where each media object is striped and stored across a group of disks where the aggregate bandwidth is at least equal to the playback rate of the object. The playback of a media stream can be seen as a sequence of periodic tasks with deadlines. These deadlines are defined by a stream playback rate which corresponds to the compression technique employed. Media streams are serviced in rounds. In each service round, data is retrieved, in parallel, from a subset of disks in the group where the object is stored across. This technique is called batched retrieval. The data retrieved is temporarily kept in a buffer. It is taken fro m the buffer and is delivered to the viewer according to the scheduled deadlines.
We have studied the effect as well as tradeoffs of the storage and retrieval parameters on number of playback sessions that the server model can support simultaneously in real-time.
According to the fact that media objects are not equally popular, one way to improve the capacity of the server is to reduce disk access cost by caching popular objects in memory. We are looking into this issue as well as to extend the server model to support streams of different playback rates and durations.