time issues and ZFS

[Daniel Braniss]

> On Jan 22, 2013, at 2:40 AM, Adrian Chadd <adrian at freebsd.org> wrote:>
> > On Jan 21, 2013, at 4:33 AM, Daniel Braniss <danny at cs.huji.ac.il>  wrote:
> > 
> >> host: DELL PowerEdge R710, 16GB, 
> 
> I administer a Dell PowerEdge R710 and I've been seeing the exact same =thing. It's currently running FreeBSD 9.0-STABLE #0 r236355. It has a =ZFS pool which sees moderate load most of the time but can be very high =at times (when certain scripts run, etc.). I hadn't previously =correlated the issue with ZFS load but that is very possible.> > I set a cron job to restart ntpd when it dies (because the time =difference exceeds the sanity check). The cron job runs "every 20 =minutes", but that varies greatly when the system stops counting. The =time offset from ntpdate (which the script runs before restarting ntpd) =varies a lot, but always in increments of 300 seconds. I've seen =everything from 1200 to 23100. (Yes, that's 23 thousand seconds aka 6 =hours 25 minutes that the system wasn't keeping time for.)
> 
> Sysctl kern.timecounter.hardware defaults to HPET. I experimented with =setting it to ACPI-fast but the issue persisted so I put it back.
> kern.timecounter.choice: TSC-low(-100) ACPI-fast(900) HPET(950) i8254(0) =dummy(-1000000)> > I first installed the box with an older 9.0-STABLE and this issue was =not present. I have been tracking -STABLE on it (albeit irregularly) so =I'm not sure when the issue came up.
> 
> 
> Have you run tests with the machdep.idle value changed, and fiddling
> 
> kern.eventtimer.periodic / kern.eventtimer.idletick ?
> 
> I would love to resolve this and am able to do some experimenting. I've =_usually_ been seeing the issue 2-3 times every 1-2 days, but I did just =make some changes:
> 	disabling ZFS compression and deduplication on all pools
> 	updated to 9.1-STABLE from yesterday (r245821)
> 
> If the issue persists I will try changing some of the sysctls above and =follow up with the result. If it goes away, I'll try to remember to =report that too.
> 
> JN
> 

set kern.eventtimer.timer=LAPIC
this solved it for me.

danny