Storage Systems for Movies-on-Demand Video Servers

1 Abstract In this paper, we evaluate storage system alternatives for movies-on-demand video servers. We begin by characterizing the movies-on-demand workload. Then we study disk farms in which one movie is stored per disk. This is a simple scheme, but it wastes substantial disk bandwidth, since disks holding less popular movies are under-utilized; also, good performance requires that movies be replicated to reeect the user request pattern. Next, we examine disk farms in which movies are striped across disks, and nd that striped video servers ooer close to full utilization of the disks by achieving better load balancing. Finally, we evaluate the use of storage hierarchies for video service that include a tertiary library along with a disk farm. Unfortunately, we show that the performance of neither magnetic tape libraries nor optical disk jukeboxes as part of a storage hierarchy is adequate to service the predicted distribution of movie accesses. We suggest changes to tertiary libraries that would make them better-suited to these applications. 2 Introduction In this paper, we discuss the design of storage systems to support movies-on-demand video service. Video-on-demand service will allow customers to request that certain movies or television shows be played on their television sets or computers immediately, giving individuals direct access to the materials stored in a video server. We examine two options for video server storage systems: collections of magnetic disks and storage hierarchies composed of magnetic disk farms and tertiary (optical disk or magnetic tape) libraries. We nd that disk systems perform best on the most important cost/performance metric for video servers: the cost per video stream. First, we study disk farms in which one movie is stored per disk. This is a simple scheme, but it wastes substantial disk bandwidth, since disks holding less popular movies are under-utilized; also, good performance requires that movies be replicated to reeect the user request pattern. Next, we examine disk farms in which movies are striped or interleaved across disks, and nd that striped video servers ooer close to full utilization of the disks by achieving better load balancing. Finally, we evaluate the use of storage hierarchies that include a tertiary library along with a disk farm. Unfortunately, we show that the performance of neither magnetic tape libraries nor optical disk jukeboxes as part of a storage hierarchy is adequate to service the predicted distribution of movie accesses. We evaluate improvements in tape libraries and …