Re: 7.0-Beta 3: zfs makes system reboot

Hello,

replying my own message is a funny thing ;-)
This message has more informational character than it is a error report.

 > Michael Rebele <m.rebele_at_web.de> wrote:
 > Von: m.rebele_at_web.de
 > Gesendet: 03.12.07 15:54:59
 > An: freebsd-current_at_freebsd.org
 > Betreff: Re: 7.0-Beta 3: zfs makes system reboot
 >
 > Alexandre Biancalana wrote:
 > > On Nov 30, 2007 2:27 PM, Michael Rebele <m.rebele_at_web.de> wrote:
 > >> 4. The applied kernel settings
 > >> kern.maxvnodes="400000"
 > >> vm.kmem_size_max="512M"
 > >> vm.kmem_size="512M"
 > >>
 > >> 5. Output from zpool
 > >> [root_at_zfs /root]# zpool status
 > >>    pool: tank
 > >>   state: ONLINE
 > >>   scrub: none requested
 > >> config:
 > >>
 > >>          NAME        STATE     READ WRITE CKSUM
 > >>          tank        ONLINE       0     0     0
 > >>            ad4s1g    ONLINE       0     0     0
 > >>
 > >> errors: No known data errors
 > >>
 > >> [kmem_map too small error]
 > >
 > > Have you tried this patch
 > > http://people.freebsd.org/~pjd/patches/vm_kern.c.2.patch ?
 >

I've applied the mentioned patch to the Beta3. Now the iozone-Benchmark 
runs through. Fine - be warned, on my machine it took about 30 hours (a 
3GHz DualCore and a 160GB SATA HD w. 10000rpm).
After the first successful run on a ZFS since my first try i tried the 
next step.
Three parallel iozone (for your reference, here's again the setup: 
iozone -R -a -z -b filez-512M.wks -g 4G -f testile) runs on the 
mentioned machine. Nearly 5 days everything went fine, but then the 
system made a reboot. Unfortunately there's no log and the reboot 
happened in the night. Though, i don't really know the reason for this. 
I guess, the kmem_map error is the cause, because the symptoms are the 
same as on the BETA3-system before the applied patch. The last output i 
have from the benchmark, showed that the file size were in the 4GB area 
with a reclen in the 8192 area (well, short before iozone should finish).
The problem is to track down the stuff, as it may took quite long until 
the error occurs - or lets say it better, even on a quite fast machine, 
iozone is quite slow. But the problem is not the CPU, it's the HD. Maybe 
somebody with access to a fast HD-Array can investigate this again.

My conclusions to my tests:
1. You should really apply the mentioned patch if you plan to use ZFS on 
  your Box for more than just testing and experimenting (well, ZFS is 
marked as that, though - you're warned). The memory/kernel parameter 
tuning recommendations helped me not really.
2. With the applied patch, ZFS seems quite robust for the average use. 
Maybe, there's a problem with a bigger load under some (rare?) 
circumstances.
3. The patch should find his way to the 'regular' kernel sources (or is 
it even in BETA4?).

A big thanks to Pawel Jakub Dawidek for his great job.

Michael

-- 
Die Erde ist die Irrenanstalt des Universums.

Public Key:
http://sks.keyserver.penguin.de:11371/pks/lookup?op=get&search=0x5D0A2BC3CEB
3F472