Clustering Active Disk Data To Improve Disk Performance

Abstract:

We show that clustering the active data of a file system in the center of the disk is a viable and effective means of improving system I/O performance. We demonstrate that it can reduce the average seek delay and in the presence of disk queues it can also reduce the average rotational delay. We also present
experimental results which show that file access patterns are strongly skewed, and that file activity levels are relatively stable over time, making them a good predictor of future file activity levels. Using simulations, we investigate two techniques for reorganizing the disk data, and we measure sensitivity to imperfect predictions of future file activity due to drifting file activity levels. We demonstrate significant performance improvements over a full spectrum of file system use, from lightly loaded systems through
heavily loaded systems with large disk queues, with performance benefits increasing as the system load increases.