An evaluation of multiple-disk I/O systems

Alternative ways of configuring an I/O subsystem with multiple disks to improve the I/O performance are considered. Specifically, the author consider disk synchronization, data declustering/disk striping, and a combination of both these approaches. They evaluate many different organizations that have not been considered before. The effects of block size and other parameters of the system are examined. Two different workloads are considered for the evaluation: a file/transaction system workload and a scientific applications workload. Through simulations it is shown that synchronized organizations perform better than other organizations at very low request rates; that there is a tradeoff in the amount of declustering/synchronization to be used in a system; and that systems with higher parallelism in reading a file perform better in a scientific workload.<
>