[OpenIndiana-discuss] Mitigating the performance impact of scrub

[Richard Elling]

On Sep 27, 2012, at 5:15 PM, Reginald Beardsley <pulaskite at yahoo.com> wrote:

> --- On Thu, 9/27/12, Richard Elling <richard.elling at richardelling.com> wrote:
>>> 
>>> zfs_scrub_delay = 100
>> 
>> a bit extreme, but probably ok
>> 
>>> zfs_scan_idle = 1000
>> 
>> no, you'll want to make this smaller.
> 
> OK, Thanks.
> 
>> Weekly scrubs are a bit extreme. Is your hardware unreliable
>> enough that you need
>> to scrub weekly? 
> 
> No reliability problems except for a 3 TB drive scratch drive that failed.  I only discovered that when I was forced to reboot after a long power failure. Both Sol 10 U8 and OI 151 have endured several power losses due to bad UPS  w/o a hiccup.

It is not likely a scrub will detect this. scrubs are mostly good for detecting
media problems and phantom writes. There are a whole bunch of other
failure modes (eg spinup) that scrubs won't exercise.

> This is new territory. I've never had the ability to check the disk integrity before.  So what is a sensible scrub schedule?  Monthly?  System runs 24x7 unless I lose power or a particularly big thunderstorm is coming through in which case I'll shut down and disconnect the cables.  It's on a UPS, but I'm on a hill in a bunch of tall trees, so it can get pretty dramatic.  A few days ago I lost 2 fluorescent light fixtures in my shop and Mom lost a cordless phone. 

Add lightning to the list of things scrub doesn't exercise :-)

The scrub interval is determined by an acceptable risk of checking for
errors that scrub exercises. Scrubs will never offer 100% no-risk.
 -- richard

--
illumos Day & ZFS Day, Oct 1-2, 2012 San Fransisco 
www.zfsday.com
Richard.Elling at RichardElling.com
+1-760-896-4422