Re: CURRENT: slow like crap! ZFS scrubbing and ports update > 25 min

On Wed, Mar 22, 2017 at 10:25:24PM +0100, O. Hartmann wrote:

> Am Wed, 22 Mar 2017 21:10:51 +0100
> Michael Gmelin <freebsd_at_grem.de> schrieb:
> 
> > > On 22 Mar 2017, at 21:02, O. Hartmann <ohartmann_at_walstatt.org> wrote:
> > > 
> > > CURRENT (FreeBSD 12.0-CURRENT #82 r315720: Wed Mar 22 18:49:28 CET 2017 amd64) is
> > > annoyingly slow! While scrubbing is working on my 12 GB ZFS volume,
> > > updating /usr/ports takes >25 min(!). That is an absolute record now.
> > > 
> > > I do an almost  daily update of world and ports tree and have periodic scrubbing ZFS
> > > volumes every 35 days, as it is defined in /etc/defaults. Prts tree hasn't grown much,
> > > the content of the ZFS volume hasn't changed much (~ 100 GB, its fill is about 4 TB
> > > now) and this is now for ~ 2 years constant. 
> > > 
> > > I've experienced before that while scrubbing the ZFS volume, some operations, even the
> > > update of /usr/ports which resides on that ZFS RAIDZ volume, takes a bit longer than
> > > usual - but never that long like now!
> > > 
> > > Another box is quite unusable while it is scrubbing and it has been usable times
> > > before. The change is dramatic ...
> > >   
> > 
> > What do "zpool list", "gstat" and "zpool status" show?
> > 
> > 
> > 
> zpool list:
> 
> NAME       SIZE  ALLOC   FREE  EXPANDSZ   FRAG    CAP  DEDUP  HEALTH  ALTROOT
> TANK00  10.9T  5.45T  5.42T         -     7%    50%  1.58x  ONLINE  -
> 
> Deduplication is off right now, I had one ZFS filesystem with dedup enabled
> 
> gstat: not shown here, but the drives comprise the volume (4x 3 TB) show 100% busy each,
> but one drive is always a bit off (by 10% lower) and this drive is walking through all
> four drives ada2, ada3, ada4 and ada5. Nothing unusual in that situation. But the
> throughput is incredible low, for example ada4:
> 
>  L(q)  ops/s    r/s   kBps   ms/r    w/s   kBps   ms/w   %busy Name
>  2    174    174   1307   11.4      0      0    0.0   99.4| ada4
> 
> kBps (kilo Bits per second I presume) are peaking at ~ 4800 - 5000. On another bos, this
> is ~ 20x higher! Most time, kBps r and w stay at ~ 500 -600.

kilo Bytes.
174 rps is normal for general 7200 RPM disk. Transfer too low by every
request is about 1307/174 = ~8 KB. Don't know root cause of this. I am
see raid-z of 4 disk, 8*3 = ~24KB per record. May be compession enable
and zfs use 128KB record size? For this case this is expected
performance. Use 1MB and higher record size.