After booting with "nosmp", I re-ran my tests. XFS is still the winner, but the
degree of its benefit from SMP is clear. Without SMP, the benchmark ran
slightly better with 5 threads (229 MB/sec under "noop"), although the
difference between "noop" and "deadline" was very slim. With SMP, the benchmark
ran noticeably better with 20 threads (378 vs. 404 MB/sec).
JFS took a small hit from the lack of SMP on 5 threads. Under 20 threads of
load, it still suffered badly, stuck between 41.5 and 45.5 MB/sec throughput.
ext3 is seems less predictable. It benefits well from SMP, but high I/O load
Mounting with "sync" showed a 21% degradation for XFS, but it still
out-performed both JFS and ext3, with or without SMP. That is, JFS and ext3
lost nearly all SMP benefit with the "sync" mount option.
On my system, XFS is the clear winner. I intend to convert my root partition to
XFS with an external journal soon, probably tomorrow night.
I am also considering posting my test observations to my website, with a story
submission to LXer.com.
Again, I have saved output from all dbench runs, if anyone wishes to review
"What better place to find oneself than
on the streets of one's home village?"
--Capt. Jean-Luc Picard, "Family"