Re: ZFS/RAIDZ and SAMBA: abyssimal performance

Fleuriot Damien <ml_at_my.gd> wrote:

> 
> On Jan 4, 2013, at 4:19 PM, O. Hartmann <ohartman_at_zedat.fu-berlin.de> wrote:
> 
> > Am 01/04/13 15:45, schrieb Garrett Cooper:
> >> On Fri, Jan 4, 2013 at 6:06 AM, Fleuriot Damien <ml_at_my.gd> wrote:
> >> 
> >> ...
> >> 
> >>> And this is under [global] in /usr/local/etc/smb.conf:
> >>>   min receivefile size = 16384
> >>>   aio read size = 16384
> >>>   aio write size = 16384
> >>>   aio write behind = yes
> >> 
> >> These are still pretty low, depending on what your networking/disk
> >> setup is like; my important performance settings are:
> >> 
> >>        socket options = SO_RCVBUF=64240 SO_SNDBUF=64240 TCP_NODELAY
> >> IPTOS_LOWDELAY IPTOS_THROUGHPUT
> >>        write cache size = 65536
> >>        aio read size = 65536
> >>        aio write size = 65536
> >>        directory name cache size = 0

> > Well, now I have peak values ~ 120 MB/s when copying. I applied Fleuriot
> > Damien's values to /boot/loader.conf and yours to the smb.conf.
> > Somewhere in the handbook this should be documented! it is to much
> > efford to get SAMBA working properly with ZFS, if the tricks and
> > problems are so widespread over several architectural aspects of the system.
> > 
> > It could save a lot of time for adminsitartors and those which try
> > FreeBSD as a serving system instead of Linux.
> > 
> > Just for the record. I feel a bit confused about all the tricks and
> > tweak now "published" for ZFS, its magic L2ARC, the kernel_vmem wizzardy
> > thingis. The ZFS Wiki seems to be a bit outdated and confusing, it would
> > be a great deal if all these things could be lined up a s a primer with
> > a bit more explanations than "put this number there".

> The problem, Oliver, is that these values are system dependant.

While I agree that the values are system dependant the purpose of
the tunables could still be documented together with a description
of how to properly test that they have any effect at all and that
it's an improvement compared to the defaults.

Scarce ZFS tuning documentation is also a problem upstream which
probably doesn't help.

Fabian