RIO

Abstract

A multimedia object server must be ready to handle a variety of media object types (video, audio, image, 3D interactive, etc.) as well as non real-time workload. Even when a homogeneous set of object types are maintained in the store (e.g., all videos) the storage system workload is generally quite variable due to the need to provide for example, VCR functionality, multiple playout rates, different resolution levels for the same objects, etc. Attempting to carefully layout data and optimally schedule delivery to meet just-in-time delivery constraints is very difficult in the face of this heterogeneous workload. Our approach to the unpredictability of the I/O workload is to randomize the allocation of disk blocks. This turns all workloads into the same uniformly random access pattern and thus gives one problem to deal with. The main disadvantage of this approach is that statistical variation can result in short term imbalances in disk utilization which in turn, cause large variances in latencies. Our approach to this problem is to introduce limited redundancy and asynchronous scheduling for short term load balancing.This approach is being implemented in the RIO (Random I/O) multimedia object server. The RIO multimedia object server provides applications a guaranteed rate of storage access with bounded delay even at very high ( > 90% ) disk utilization.