Sysbench read/write speed
Another synthetic benchmark that measure read/write speed is Sysbench. Lets begin with sequential read/write tests:
Sequential performance are mostly influenced by hard disk's linear read/write speed,so the results are quite similar here. However, we can see that xfs seems to take a small lead, especially in the case of synchronized operations.
We can see that ext4 is the slower in random read, while ext3 and btrfs are very fast. However, random read results seems absolutely too high to realistically depict filesystems performance: maybe this test is not using a real random pattern, or it is inflated by hard-drive and o.s. read-ahead settings. Whichever is the cause, the synchronized write test seems to be not affected: its results are in-line with what this old 80 GB hard disk can deliver. We see that when not synching each request, btrfs and ext3 are way faster then ext4 and xfs. When issuing synchronized write operations btrfs become the slower of the lot, while ext3 and xfs are noticeably faster, with ext4 in between.
Now, lets see filesystems performance with direct access mode. Sequential tests first:
Sequential speed results are mostly unchanged. Maybe random results will give us some surprise?
Uhm, yes: btrfs read speed grow to inexplicable levels. Something is clearly wrong with Sysbench random read test. However, random write test reports consistent results, and it tell us that little has changed from non-direct access mode.