Cluster Placement: a Data Placement Scheme for a Mass Storage System of Video-on-demand Servers

The retrieval process of a continuous media is considered as a periodic request which each subrequest retrieves a block of data. The next data block has to be ready before the display of the current data block is completed. The size of data block, called buuer size, along with the access time of the data block then determine the number of such requests that can be supported. This implies that to reduce access time, it can either decrease the required buuer size and/or increase the number of retrieval processes. This paper addresses the issue of reducing access time from data placement perspective. We propose a data placement scheme called Cluster Placement for a mass storage system of Video-On-Demand (VOD) servers. The cluster placement divides disk cylinders into equal-size clusters and stores video data across clusters. Retrieval processes are synchronized such that a batch of subrequests are within a cluster range. However, the cluster accessed by the current batch of subrequests may diierent from the one accessed by the next batch. The range of disk seek operations for a batch of subrequests can then be limited within a cluster range compared to the whole disk area in the normal case. Consequently, reduces the disk access time. We identify several design issues associated with the cluster placement scheme; one important issue is the relationship between underlying disk scheduling policies with video placement schemes. We illustrated the eeectiveness of the proposed cluster placement by applying NTSC video quality into an array of eight disks under FIFO, C-SCAN, and GSS disk scheduling schemes. The analytic result shows that the proposed cluster placement can reduce the required buuer size from 3% to 40% when compared to a contiguous placement. We also conducted an experiment on SGI Onyx system which connected to a level-3 RAID (Redundant Array of Inexpensive Disks). The result shows that the proposed cluster placement can reduce buuer size up to 25% or support one more retrieval process when compared to the contiguous placement.