A file assignment strategy for parallel I/O system with minimum I/O contention probability

Abstract

Many solutions have been proposed to tackle the problem of assigning files in a parallel I/O system. The primary objective of the existing solutions is either to balance the load among disks or to minimize the service time variance at each disk, whereas the dynamic characteristics of the file requests which would access these files are ignored. The studies on the dynamic I/O behaviors of applications show that the file requests targeted on the different popular files situated in the same disk may temporally compete with each another for the same disk. Consequently, the performance gained from the parallelism of multiple disks is degraded because this type of I/O contention turns the parallel I/O into sequential one. Hence, how to minimize the I/O contention among the file requests should become one of the new objectives which the file assignment strategy should take into consideration. In order to address this issue, this study proposes a new static file assignment algorithm named MinCP for parallel I/O system. Through assigning files sorted in their access rates onto multiple disks in round-robin fashion, the MinCP aims to minimize the I/O contention probability among file requests, thereby optimizing the mean response time of these requests. The experiment results show that the MinCP achieves optimal performance on mean response time among the existing schemes for comparison. © 2011 Springer-Verlag.