After many trials (and errors), here are some facts: host: DELL PowerEdge R710, 16GB, mfi0: <Dell PERC H700 Integrated> mfid0: 14305280MB (29297213440 sectors) RAID volume 'r5' is optimal mfi1: <Dell PERC 6> mfid1: 12393472MB (25381830656 sectors) RAID volume 'Virtual Disk 0' is optimal we have NO problems with FreeBSD-8.3-STABLE, but with 9.1-STABLE, the real-time clock slows down when doing some zfs stuff like send|receive, typing 'date' when less that 1000s went by seems to crorrect the problem, ntpd kicks in and on track again. I have a cron job just logging date every 5 minutes, and the loghost sees: |-- local time on loghost | time on problematic host Jan 20 19:56:19 store-02.cs.huji.ac.il Jan 20 19:56:19 danny: Sun Jan 20 19:56:19 IST 2013 -- ok Jan 20 20:15:00 store-02.cs.huji.ac.il Jan 20 20:15:00 danny: Sun Jan 20 20:15:00 IST 2013 -- ok Jan 20 21:30:00 store-02.cs.huji.ac.il Jan 20 20:21:06 danny: Sun Jan 20 20:21:06 IST 2013 -- off by 1:09 Jan 20 21:33:53 store-02.cs.huji.ac.il Jan 20 20:25:00 danny: Sun Jan 20 20:25:00 IST 2013 -- off by 1:08 Jan 20 21:38:54 store-02.cs.huji.ac.il Jan 20 20:30:00 danny: Sun Jan 20 20:30:00 IST 2013 -- off by 1:09 ... Jan 20 22:03:54 store-02.cs.huji.ac.il Jan 20 20:55:00 danny: Sun Jan 20 20:55:00 IST 2013 -- diff is now constant .. Jan 20 22:04:13 store-02.cs.huji.ac.il Jan 20 20:55:19 ntpd[1848]: time correction of 4134 seconds exceeds sanity limit (1000); set clock manually to the correct UTC time. ... Jan 20 22:58:53 store-02.cs.huji.ac.il Jan 20 21:50:00 danny: Sun Jan 20 21:50:00 IST 2013 strangely, when running 8.3, ACPI-fast is chosen: kern.timecounter.choice: TSC(-100) HPET(900) ACPI-fast(1000) i8254(0) dummy(-1000000) but with 9.1 TSC-low gets chosen: kern.timecounter.choice: TSC-low(1000) ACPI-fast(900) HPET(950) i8254(0) dummy(-1000000) so I did sysctl kern.timecounter.hardware=ACPI-fast, but the same happens - unless it can't be changed after boot. I realy need help here! thanks, danny
Hi, Try experimenting with kern.eventtimer.periodic and kern.eventtimer.idletick. If this fixes it for you, please file a PR with all the relevant details. Thanks! Adrian On 21 January 2013 03:33, Daniel Braniss <danny at cs.huji.ac.il> wrote:> After many trials (and errors), here are some facts: > > host: DELL PowerEdge R710, 16GB, > mfi0: <Dell PERC H700 Integrated> > mfid0: 14305280MB (29297213440 sectors) RAID volume 'r5' is optimal > mfi1: <Dell PERC 6> > mfid1: 12393472MB (25381830656 sectors) RAID volume 'Virtual Disk 0' is > optimal > > we have NO problems with FreeBSD-8.3-STABLE, but with 9.1-STABLE, the real-time > clock slows down when doing some zfs stuff like send|receive, typing 'date' > when less that 1000s went by seems to crorrect the problem, > ntpd kicks in and on track again. > > I have a cron job just logging date every 5 minutes, and the loghost sees: > > |-- local time on loghost | time on problematic host > Jan 20 19:56:19 store-02.cs.huji.ac.il Jan 20 19:56:19 danny: Sun Jan 20 > 19:56:19 IST 2013 -- ok > Jan 20 20:15:00 store-02.cs.huji.ac.il Jan 20 20:15:00 danny: Sun Jan 20 > 20:15:00 IST 2013 -- ok > Jan 20 21:30:00 store-02.cs.huji.ac.il Jan 20 20:21:06 danny: Sun Jan 20 > 20:21:06 IST 2013 -- off by 1:09 > Jan 20 21:33:53 store-02.cs.huji.ac.il Jan 20 20:25:00 danny: Sun Jan 20 > 20:25:00 IST 2013 -- off by 1:08 > Jan 20 21:38:54 store-02.cs.huji.ac.il Jan 20 20:30:00 danny: Sun Jan 20 > 20:30:00 IST 2013 -- off by 1:09 > ... > Jan 20 22:03:54 store-02.cs.huji.ac.il Jan 20 20:55:00 danny: Sun Jan 20 > 20:55:00 IST 2013 -- diff is now constant > .. > Jan 20 22:04:13 store-02.cs.huji.ac.il Jan 20 20:55:19 ntpd[1848]: time > correction of 4134 seconds exceeds sanity limit (1000); set clock manually to > the correct UTC time. > ... > Jan 20 22:58:53 store-02.cs.huji.ac.il Jan 20 21:50:00 danny: Sun Jan 20 > 21:50:00 IST 2013 > > > strangely, when running 8.3, ACPI-fast is chosen: > kern.timecounter.choice: TSC(-100) HPET(900) ACPI-fast(1000) i8254(0) > dummy(-1000000) > but with 9.1 TSC-low gets chosen: > kern.timecounter.choice: TSC-low(1000) ACPI-fast(900) HPET(950) i8254(0) > dummy(-1000000) > > so I did sysctl kern.timecounter.hardware=ACPI-fast, but the same happens - > unless it can't be changed after boot. > > I realy need help here! > > thanks, > danny > > > > _______________________________________________ > freebsd-stable at freebsd.org mailing list > http://lists.freebsd.org/mailman/listinfo/freebsd-stable > To unsubscribe, send any mail to "freebsd-stable-unsubscribe at freebsd.org"
On Mon, 2013-01-21 at 13:33 +0200, Daniel Braniss wrote:> After many trials (and errors), here are some facts: > > host: DELL PowerEdge R710, 16GB, > mfi0: <Dell PERC H700 Integrated> > mfid0: 14305280MB (29297213440 sectors) RAID volume 'r5' is optimal > mfi1: <Dell PERC 6> > mfid1: 12393472MB (25381830656 sectors) RAID volume 'Virtual Disk 0' is > optimal > > we have NO problems with FreeBSD-8.3-STABLE, but with 9.1-STABLE, the real-time > clock slows down when doing some zfs stuff like send|receive, typing 'date' > when less that 1000s went by seems to crorrect the problem, > ntpd kicks in and on track again. > > I have a cron job just logging date every 5 minutes, and the loghost sees: > > |-- local time on loghost | time on problematic host > Jan 20 19:56:19 store-02.cs.huji.ac.il Jan 20 19:56:19 danny: Sun Jan 20 > 19:56:19 IST 2013 -- ok > Jan 20 20:15:00 store-02.cs.huji.ac.il Jan 20 20:15:00 danny: Sun Jan 20 > 20:15:00 IST 2013 -- ok > Jan 20 21:30:00 store-02.cs.huji.ac.il Jan 20 20:21:06 danny: Sun Jan 20 > 20:21:06 IST 2013 -- off by 1:09 > Jan 20 21:33:53 store-02.cs.huji.ac.il Jan 20 20:25:00 danny: Sun Jan 20 > 20:25:00 IST 2013 -- off by 1:08 > Jan 20 21:38:54 store-02.cs.huji.ac.il Jan 20 20:30:00 danny: Sun Jan 20 > 20:30:00 IST 2013 -- off by 1:09 > ... > Jan 20 22:03:54 store-02.cs.huji.ac.il Jan 20 20:55:00 danny: Sun Jan 20 > 20:55:00 IST 2013 -- diff is now constant > .. > Jan 20 22:04:13 store-02.cs.huji.ac.il Jan 20 20:55:19 ntpd[1848]: time > correction of 4134 seconds exceeds sanity limit (1000); set clock manually to > the correct UTC time. > ... > Jan 20 22:58:53 store-02.cs.huji.ac.il Jan 20 21:50:00 danny: Sun Jan 20 > 21:50:00 IST 2013 > > > strangely, when running 8.3, ACPI-fast is chosen: > kern.timecounter.choice: TSC(-100) HPET(900) ACPI-fast(1000) i8254(0) > dummy(-1000000) > but with 9.1 TSC-low gets chosen: > kern.timecounter.choice: TSC-low(1000) ACPI-fast(900) HPET(950) i8254(0) > dummy(-1000000) > > so I did sysctl kern.timecounter.hardware=ACPI-fast, but the same happens - > unless it can't be changed after boot. > > I realy need help here! > > thanks, > dannyWhat's the output of sysctl kern.eventtimer? Does the bad behavior change if you set kern.eventimer.periodic=1? -- Ian
On Tue, Jan 22, 2013 at 7:27 AM, Julian Stecklina < jsteckli at os.inf.tu-dresden.de> wrote:> Does anyone know why the LAPIC is given a lower priority than HPET in > this case? If you have an LAPIC, it should always be prefered to HPET, > unless something is seriously wrong with it... >On many processors the lapic timer does not work correctly in states lower than C1. There are many processors that will automatically enter a "C1E" mode when the processor is idle, and in that state I have seen the lapic timer run slower than the programmed frequency, causing time to move to slowly on idle FreeBSD systems.