cost of tdb transactions on ext4 with barriers

[simo]

On Fri, 2010-02-12 at 12:44 +1100, tridge at samba.org wrote:
> Hi Rusty,
> 
> for completeness, here is are the results including the btrfs and xfs
> numbers:
> 
>    ext3 barrier=0   1.9s
>    ext3 barrier=1   13.4s

Tridge,
can you try the above test with data=ordered ?
AFAIK ext4 has that option on by default.

>    ext4 barrier=0   2.5s
>    ext4 barrier=1   36.0s
> 
>    btrfs nobarrier  29.1s  
>    btrfs barrier=1  117s
> 
>    xfs   	    8.9s  

I am told xfs dumps to disk only the data related to the specific
fsync(), while ext3 (and probably ext4) dump *everything*, even
unrelated data, to disk on fsync.
Btrfs should do the right thing in time, but I assume the version youare
running is not quite prime time ready or tuned.

> (the barrier options make no difference on xfs as far as I can see).
> 
> If ext4 and btrfs are the future of Linux filesystems then tdb/Samba
> are in real trouble!

I don't think the current data you have btrfs really represent the
performance of btrfs when it will be ready. However I am told that
btrfs's COW may be a problem, and lots of data rewrites amy be a worst
case scenario for btrfs.

Can you test btrfs with nodatacow ?

Simo.

-- 
Simo Sorce
Samba Team GPL Compliance Officer <simo at samba.org>
Principal Software Engineer at Red Hat, Inc. <simo at redhat.com>