WC invcstigatc design principIcs for placing striped delayscnsitivc data on a number of disks in a distributed environment, The cost formulas allow us to calculate the maximum number of users that can be supported by n disks and the minimum number of disks needed to support k users, as well 118 to study the impact of other performance tuning options. Next, WC examine the problem of optimal placement for striped data, WC show that for fixed probabilities of accessing the delay-sensitive objects, partitioning the set of disks is always bcttcr than striping in all of the disks. Then, given a number it of disks and T distinct delay-sensitive objects with probubilitics of access p~,pz,. . . ,pr, that must be striped across T different disk partitions (i.e., non-overlapping subsets of the n disks), WC use the Majorization theory and the theory of Schur functions in order to find what is the optimal number of disks that must be allocated to each partition. WC analyze the problem of grouping the more and less popular deluy-sensitive objects together in partitions when the partitions arc less than the objects, so that the number of supported USCPS is maximized. Finally, we present the tradeoff of striping on all the disks versus partitioning the set of the dislm when the access probabilities of the delay-sensitive objects change with time.
Unfortunately, ACM prohibits us from displaying non-influential references for this paper.
To see the full reference list, please visit http://dl.acm.org/citation.cfm?id=275496.