In an effort to meet the increasing throughput demands on the SunOS file system made both by applications and higher performance hardware, several optimization paths were examined. The principal constraints were that the on−disk file system format remain the same and that whatever changes were necessary not be user−visible. The solution arrived at was to approximate the behavior of extent based file systems by grouping I/O operations into clusters instead of dealing in individual blocks. A single clustered I/O may take the place of 15−30 block I/Os, resulting in a factor of two increased sequential performance increase. The changes described were restricted to a small portion of the file system code; no user−visible changes were necessary and the on-disk format was not altered.

In an effort to meet the increasing throughput demands on the SunOS file system made both by applications and higher performance hardware, several optimization paths were examined. The principal constraints were that the on-disk file system format remain the same and that whatever changes were necessary not be user-visible. The solution arrived at was to approximate the behavior of extent based file systems by grouping I/O operations into clusters instead of dealing in individual blocks. A single clustered I/O may take the place of 15-30 block I/Os, resulting in a factor of two increased sequential performance increase. The changes described were restricted to a small portion of the file system code; no user-visible changes were necessary and the on-disk format was not altered. Introduction File systems are a common place to find performance problems. The original UNIX file system [Thompson] is elegant in its simplicity: it has a single block size and a simple list based allo...