Variable bit rate MPEG video storage on parallel disk arrays

We discuss issues related to the storage and retrieval of variable bit rate (VBR) MPEG video data on parallel disk arrays for multiple simultaneous users. We store data blocks corresponding to constant real-time playback duration. Because of the VBR characteristics and the constant bandwidth read channel, disk overload occurs at peak usage periods. Our strategy to deal with disk overload is to temporarily stop service to low-priority users. To do so, we propose a number of classes of service corresponding to various probabilities of loss, during which time service is suspended. Thus, a major problem is estimating probability of loss or suspension so that a video server system can guarantee service with actual loss probabilities that do not exceed the specified thresholds at the call setup time. We propose three techniques for computing loss probabilities: histogram convolution, Central Limit Theorem, and Cramer's rule. We present an efficient placement strategy and an optimal admissions control strategy that guarantees loss probability thresholds while maximizing the number of requests that can be satisfied at all qualities. We extend these ideas to encompass scalability by allowing appropriate frames in the MPEG-encoded data to be dropped without fully suspending service to any one user. This allows the system to make intelligent choices in gracefully degrading the request data rates during periods of peak usage. Finally, we test the data placement and admissions control algorithms with a discrete event disk array system simulator and show that the results are in accordance with theoretical predictions.<>