ZFS (ARC) performance regression in r321610

[Yamagi Burmeister]

Hi,
I just want a give a short conclusion for this problem.

On Mon, 27 Aug 2018 17:19:10 +0300
Lev Serebryakov <lev at FreeBSD.org> wrote:

> On 27.08.2018 16:47, Yamagi Burmeister wrote:
> 
> > With this program I was able to bisect the source and identify commit
> > r321610 (MFV r318946: 8021 ARC buf data scatter-ization) as the culprit:
> > https://svnweb.freebsd.org/base?view=revision&revision=321610
>  You could try to set "zfs_abd_scatter_enabled = 0" via kgdb and repeat
> tests to be sure, that this code is a problem.

After about 2 weeks with "zfs_abd_scatter_enabled" set to 0 (in form of
r338365 merged back to 11.2-RELEASE and vfs.zfs.abd_scatter_enabled set
to 0) I can confirm that switching of scattered allocations helps, my
performance problems are gone. The system is running fine and zfs-stats
looks good.

To be clear: Switching off scattered allocations has a very positive
effect in this and only this workload. On our other ~100 hosts - some of
them MySQL installations with other workload patterns - the scattered
allocations are a benefit or at least don't hurt.

Regards,
Yamagi

-- 
Homepage: https://www.yamagi.org
Github:   https://github.com/yamagi
GPG:      0x1D502515
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 833 bytes
Desc: not available
URL: <http://lists.freebsd.org/pipermail/freebsd-fs/attachments/20180907/87710d14/attachment.sig>