[Linux-ia64] NTP broken?

[Linux-ia64] NTP broken?

[Peter Chubb]

NTP on Linux 2.4.17/IA64 appears broken.   ntpd never seems to be able
to pull the PLL into sync, even after using ntpdate to set the clock
to the correct server time.

What are other peoples' experiences with NTP on IA64?  

Peter C

Re: [Linux-ia64] NTP broken?

[Andreas Schwab]

Peter Chubb <peter@chubb.wattle.id.au> writes:

|> NTP on Linux 2.4.17/IA64 appears broken.   ntpd never seems to be able
|> to pull the PLL into sync, even after using ntpdate to set the clock
|> to the correct server time.
|> 
|> What are other peoples' experiences with NTP on IA64?  

It appears to be working correctly on all our machines, with various
kernel versions.

Andreas.

-- 
Andreas Schwab, SuSE Labs, schwab@suse.de
SuSE GmbH, Deutschherrnstr. 15-19, D-90429 Nürnberg
Key fingerprint = 58CA 54C7 6D53 942B 1756  01D3 44D5 214B 8276 4ED5
"And now for something completely different."

Re: [Linux-ia64] NTP broken?

[David Mosberger]

>>>>> On Wed, 22 May 2002 13:08:08 +1000, Peter Chubb <peter@chubb.wattle.id.au> said:

  Peter> NTP on Linux 2.4.17/IA64 appears broken.  ntpd never seems to
  Peter> be able to pull the PLL into sync, even after using ntpdate
  Peter> to set the clock to the correct server time.

  Peter> What are other peoples' experiences with NTP on IA64?

I'm not aware of any problems with NTP keeping in sync with realtime.
However, I do think the drift on the kernel's time-of-day (without NTP
synchronization) may be bigger than it could be.  This could be due to
rounding errors, due to the CPU clock not being accurate enough, or
something else entirely.  Some time ago, I did look over the code and
didn't find any obvious rounding errors, but that doesn't mean they
don't exist.  Actually, one experiment that might be worth doing is to
set HZ to 100.  This will make maintaining the time-of-day slightly
easier and since it also happens to be the freq used by x86, makes it
almost certain that there are no other subtle issues.

	--david

RE: [Linux-ia64] NTP broken?

[Luck, Tony]

I compared a HZ\x100 kernel with a HZ\x1024 kernel on a BigSur ... it
made no obvious difference, both kernels lost time at about 38.5
milliseconds/minute relative to a machine running NTP.

-Tony

-----Original Message-----
From: David Mosberger [mailto:davidm@napali.hpl.hp.com]
Sent: Friday, May 24, 2002 7:33 PM
To: Peter Chubb
Cc: linux-ia64@linuxia64.org
Subject: Re: [Linux-ia64] NTP broken?


>>>>> On Wed, 22 May 2002 13:08:08 +1000, Peter Chubb
<peter@chubb.wattle.id.au> said:

  Peter> NTP on Linux 2.4.17/IA64 appears broken.  ntpd never seems to
  Peter> be able to pull the PLL into sync, even after using ntpdate
  Peter> to set the clock to the correct server time.

  Peter> What are other peoples' experiences with NTP on IA64?

I'm not aware of any problems with NTP keeping in sync with realtime.
However, I do think the drift on the kernel's time-of-day (without NTP
synchronization) may be bigger than it could be.  This could be due to
rounding errors, due to the CPU clock not being accurate enough, or
something else entirely.  Some time ago, I did look over the code and
didn't find any obvious rounding errors, but that doesn't mean they
don't exist.  Actually, one experiment that might be worth doing is to
set HZ to 100.  This will make maintaining the time-of-day slightly
easier and since it also happens to be the freq used by x86, makes it
almost certain that there are no other subtle issues.

	--david

_______________________________________________
Linux-IA64 mailing list
Linux-IA64@linuxia64.org
http://lists.linuxia64.org/lists/listinfo/linux-ia64

RE: [Linux-ia64] NTP broken?

[David Mosberger]

>>>>> On Tue, 28 May 2002 15:48:49 -0700, "Luck, Tony" <tony.luck@intel.com> said:

  Tony> I compared a HZ\x100 kernel with a HZ\x1024 kernel on a BigSur
  Tony> ... it made no obvious difference, both kernels lost time at
  Tony> about 38.5 milliseconds/minute relative to a machine running
  Tony> NTP.

Hmmh, interesting data.  So the rate is 633ppm.  Does anyone know what
the crystals on the BigSurs are rated at?

	--david

Re: [Linux-ia64] NTP broken?

[Peter Chubb]

>>>>> "David" = David Mosberger <davidm@napali.hpl.hp.com> writes:

>>>>> On Wed, 22 May 2002 13:08:08 +1000, Peter Chubb <peter@chubb.wattle.id.au> said:
Peter> NTP on Linux 2.4.17/IA64 appears broken.  ntpd never seems to
Peter> be able to pull the PLL into sync, even after using ntpdate to
Peter> set the clock to the correct server time.

Peter> What are other peoples' experiences with NTP on IA64?

David> I'm not aware of any problems with NTP keeping in sync with
David> realtime.  

OK, I now have two i2000s both running the same kernel, both
configured identically.  On one of them, NTP pulls into sync within a
few minutes.  The other just drifts further and further away.

One interesting `feature' is that on the one that keeps running away,
the CPU clock is detected as 718.793988 MHz not the supposedly correct
733MHz.  The boot message says:
	 CPU 0: base freq\x130.684MHz, ITC ratio\x11/2, ITC
		freqq8.764MHz
My suspicion is that this is incorrect on this machine, and it is
*really* at 733MHz --- which would explain the lack of sync (NTP just
can't slow the clock down enough) and the runaway time (733/718 is
around 1.02; which corresponds to a second a minute) 

The BIOS reports 733MHz, though.

Peter C

Re: [Linux-ia64] NTP broken?

[Peter Chubb]

>>>>> "Peter" = Peter Chubb <peter@chubb.wattle.id.au> writes:

>>>>> "David" = David Mosberger <davidm@napali.hpl.hp.com> writes:
>>>>> On Wed, 22 May 2002 13:08:08 +1000, Peter Chubb <peter@chubb.wattle.id.au> said:
Peter> One interesting `feature' is that on the one that keeps running
Peter> away, the CPU clock is detected as 718.793988 MHz not the
Peter> supposedly correct 733MHz.  

Guess what.  After the reboot that I did to find out what the BIOS
said, the kernel now reports the correct cpu clock frequency, and NTP
works.

Peter C

Re: [Linux-ia64] NTP broken?

[David Mosberger]

>>>>> On Wed, 29 May 2002 15:13:05 +1000, Peter Chubb <peter@chubb.wattle.id.au> said:

  Peter> Guess what.  After the reboot that I did to find out what the
  Peter> BIOS said, the kernel now reports the correct cpu clock
  Peter> frequency, and NTP works.

Are you running the latest firmware?  The clockfrequency is indeed
_measured_ by the firmware and there used to be occasional but large
errors in that measurement.  I thought newer firmware doesn't have
this problem any longer.  So, if you're not running the latest version
yet, it might be worth to update the firmware.

	--david

Re: [Linux-ia64] NTP broken?

[Rich Altmaier]

David, FYI, I asked our hardware designers about current oscillator
accuracy and was a bit suprised to learn that one pays a
premium price today to get 300ppm.  600ppm can happen
in lower cost units.

FYI, Rich



David Mosberger wrote:
> 
> >>>>> On Tue, 28 May 2002 15:48:49 -0700, "Luck, Tony" <tony.luck@intel.com> said:
> 
>   Tony> I compared a HZ\x100 kernel with a HZ\x1024 kernel on a BigSur
>   Tony> ... it made no obvious difference, both kernels lost time at
>   Tony> about 38.5 milliseconds/minute relative to a machine running
>   Tony> NTP.
> 
> Hmmh, interesting data.  So the rate is 633ppm.  Does anyone know what
> the crystals on the BigSurs are rated at?
> 
>         --david
> 
> _______________________________________________
> Linux-IA64 mailing list
> Linux-IA64@linuxia64.org
> http://lists.linuxia64.org/lists/listinfo/linux-ia64

Re: [Linux-ia64] NTP broken?

[David Mosberger]

>>>>> On Wed, 29 May 2002 10:18:43 -0700, Rich Altmaier <richa@engr.sgi.com> said:

  Rich> David, FYI, I asked our hardware designers about current
  Rich> oscillator accuracy and was a bit suprised to learn that one
  Rich> pays a premium price today to get 300ppm.  600ppm can happen
  Rich> in lower cost units.

Thanks for sharing that.  I just tried the obvious google search for
"crystal ppm" and most of the hits where for crystals on the order of
30-50ppm, with the occasional 100ppm.  Precision crystals seem to be
rated at <1ppm.  So, 600ppm seems a bit on the high side, but I guess
it's not totally unreasonable either.  I guess NTP is our
friend... ;-)

	--david