just another sad story in zfs tuning city

Marco van Tol marco at tols.org
Fri Sep 3 13:06:50 UTC 2010


On Fri, Sep 03, 2010 at 09:58:28PM +0900, Randy Bush wrote:
> i have a system that should move right along.  but it is as soggy as a
> tenugui on an august afternoon.
> 
> 2*amd64 2.0GHz 4GB 750GB
> 8.1-STABLE Sun Aug  8 01:05:24 UTC 2010
> 
> /dev/mirror/boota                     8.3G    909M    6.7G    12%    /
> procfs                                4.1k    4.1k      0B   100%    /proc
> tank                                  729G     28k    729G     0%    /tank
> tank/data                             729G     31k    729G     0%    /data
> tank/data/nfsen                       877G    148G    729G    17%    /data/nfsen
> tank/data/rpki                        729G    107M    729G     0%    /data/rpki
> tank/usr                              737G    7.7G    729G     1%    /usr
> tank/usr/home                         732G    2.6G    729G     0%    /usr/home
> tank/usr/usr                          732G    2.6G    729G     0%    /usr/usr
> tank/var                              730G    759M    729G     0%    /var
> tank/var/log                          730G    415M    729G     0%    /var/log
> tank/var/spool                        729G     68M    729G     0%    /var/spool
> 
> work0.psg.com:/usr/home/rancid# cat /boot/loader.conf.local
> loader_logo=beastie
> zfs_load=YES
> vm.kmem_size=4G
> vfs.zfs.arc_max=64M
> vfs.zfs.prefetch_disable=1
> geom_mirror_load=YES
> kern.maxvnodes=50000
> 
> i am tempted to just boot without the zfs memory hacks in loader conf.
> any warnings on doing so?  any other clues also gladly accepted.

In my view you limited the page cache (equivalent) for your zfs
filesystems to 64MB out of 4GB system memory.  Not sure what your I/O is
like, but that could explain the sluggishness? :-)

Sticking a wet finger in the air, and assuming /dev/mirror/boota holds a
UFS filesystem that's just used to boot from I'd say give kmem_size 1.5x
physical, arc_max 1.0x physical, and arc_min 0.5x physical memory to
have it compete with the UFS page cache.  After that applications that
run can't use more then (0.5x physical memory plus swap space).

Then again, I'm by far no zfs expert, so consider this an experiment ;)

Marco

-- 
Now watch me snatch defeat from the jaws of victory
- "Rigoletto" during a game on www.dailygammon.com


More information about the freebsd-fs mailing list