Bill,
Another case in which the engineering model in Linux and NTP are not
compatible. Neither is necessarily wrong, just different. The following
issues are known to me.
1. The Linux kernel discipline code adapted from my Alpha code of the
1990s does not account for the frequency gain at other than a 100-Hz
clock. With a 1000-Hz clock this results in serious instability. I
pointed this out some years ago and it is a trivial modification, but so
far as I know it has not been fixed.
2. The Linux adjtime mechanism inserts an extra pole in the impulse
response, presumably to speed convergence when relatively large
adjustments are made. This makes NTP unstable at the larger poll
intervals when the kernel discipline is not in use. Both the kernel and
daemon discipline loops are carefully designed according to sound
engineering principles for optimum response, but the extra pole defeats
the design.
3. The calling sequence for the ntp_gettime() system call is
incompatible with current use. As a result, access to the TAI-UTC offset
by application programs is not available.
4. As in the current instance, management of the RTC and system clock is
incompatible. This issue should be reviewed in the context of the
various models, whether the kernel or daemon discipline is in use and
whether the system is awake or sleeping.
There are probably others and they probably could be resolved to insure
a consistent model between Linux and other operating systems.
Dave
unruh wrote:
On 2010-06-23, David L. Mills <mi...@udel.edu> wrote:
Pavel,
Linux has many, many times broken the NTP model compatible with other
systems such as Solaris and FreeBSD, among others. I have no trouble
with that as long as whatever modifications are required in NTP to make
the RTC driver work remain proprietary to Linux and never leak to other
systems. I have no idea what the Linux 11-minute process is about, but
it probably conflicts with the NTP 1-hour RTC alignment.
Linux, depending on the setting of a flag in the adjtimex setup, sets
the rtc from the system time once every 11 min. . This is a disaster if
you have a procedure to discipline the rtc (eg hwclock, or chrony) and
the "sychronization " flag must be kept unset to prevent this behaviour.
On most systems the rtc is used to set the clock whent he computer is
down . Ie, the rtc in those cases CANNOT be disciplined. All you can do
is to determine the offset and drift rate of the rtc to make the use as
accurate as possible when the system comes up again. It is very hard to
detrmine the drift rate of a clocck that keeps getting reset
To then use the rtc mechanism in a VM seems to me to be overloading the
mechanism, making it hard to anything reasonable with it.
If your driver requires the Linux model, whatever modifications are
required in the base code (#ifdefs) will not be supported here and may
conflict with future developments. On the other hand, it could be, for
example, that the RTC provide a 1-second interrupt similar to the PPS
signal now. On that assumption the base code might have a feature that
supports the RTC in much the same way the NMEA driver does now. That
would be a generic solution and nicely fit the NTP.
Dave
Krejci, Pavel wrote:
Hello Dave,
------------------------------------------------------------------------
From: David L. Mills [mailto:mi...@udel.edu]
Sent: Wednesday, June 23, 2010 4:42 AM
To: Krejci, Pavel
Cc: questions@lists.ntp.org
Subject: Re: [ntp:questions] Reference clock driver for /dev/rtc
Pavel,
It's not as simple as that. Normally, ntpd uses settimeofday()
once per hour to set the system clock, which has the side effect
of setting the RTC. Obviously, you don't want that. If the RTC
refclock is enabled, that has to be disabled, so some kind of
interlock must be devised. This can be a tricky business and have
unintended consequences if something or other fails. The
interlocks with the PPS signal come to mind.
Do you mean the 11 minute mode in Linux, when the system time is
periodically written to the rtc in 11 minute intervals? This is
triggered by the synch status (time_status variable in the
kernel). I've solved this by periodically resetting this synch
status in my refclock driver.
You are correct in that the RTC has in general far better
temperature compensation than either the system clock or the
TSC/PCC counter. However, its resolution is generally far worse.
Even so, the lowpass character of the clock discipline masks this
so actual delivered system time should be quite good. Chapter 15
of my new book due in September contains an extensive discussion
on these issues.
Theoretically the worst RTC resolution is 1 second, but usually if
offers update IRQ whenever the "seconds" counter changes. And this
gives good resolution for my system. Attached is the /dev/rtc
peerstats from my qemu guest system. The clock offset keeps
under 1 milisecond which is enough for our purposes. I will check
your book when published.
Regards
Pavel
Dave
Krejci, Pavel wrote:
Hi,
well, then, do you find it useful? How should I proceed to contribute into ntpd
project?
Thanks
Pavel
-----Original Message-----
From: unruh [mailto:un...@wormhole.physics.ubc.ca]
Sent: Thursday, June 17, 2010 11:48 AM
To: questions@lists.ntp.org
Subject: Re: [ntp:questions] Reference clock driver for /dev/rtc
On 2010-06-16, Krejci, Pavel
<pavel.kre...@siemens-enterprise.com> wrote:
Hi,
-----Original Message-----
From: unruh [mailto:un...@wormhole.physics.ubc.ca]
Sent: Tuesday, June 15, 2010 7:15 PM
To: questions@lists.ntp.org
Subject: Re: [ntp:questions] Reference clock driver for /dev/rtc
On 2010-06-15, Krejci, Pavel
<pavel.kre...@siemens-enterprise.com> wrote:
Hi,
since I cannot use kvm-clock as the clock source (older
guest kernel) I am using pit as the clock source. According to my
tests this is the most stable clock source among tsc,hpet
but still
can drift. Since the qemu keeps the /dev/rtc perfectly
synchronized
with the Host's system time it is a good time source for
the ntpd on
the guest. The host itself is then sychronized via NTP with the
external time server. I don't know any other way how to read the
system time from the Host, please offer if you have some.
I do not understand. If you driver can read the rtc, it
can read the
system clock instead.
I am not reading the Host's /dev/rtc. I am reading the
Guest's /dev/rtc, which is synchronized with the Host's system clock.
OK, if that is the way your virtual system works, (Ie it
delivers the system time via /dev/rtc) then so be it. I would
say it is terrible, since it uses a predefined item ( rtc) to
deliver something totally different ( the system time of the
underlying host) rtc has numberous idiosyncracies, not oleast
being that it delivers only times with one second precision.
It also delivers an interrupt on one second boundaries, is
written by a displacement of .5 sec (Ie if you write the time
x to it, that time refers to the time of the rtc .5 sec in
the future. ) I assume that your /dev/rtc does not have all
thoese peculiarities.
And virtual systems are terrible things to use ntpd on.
ntpd cannot
control something where the clock varies by more than 500PPM, and
virtual systems, since they are shut down for variable lengths of
time by the host, tend to have terrible clocks.
Yes the clocks like hpet or tsc are drifting very very much
and the ntpd cannot improve the resulting stability. But the
pit keeps quite well. With additional ntpd the resulting long
period clock stability is good - no exact measures done yet...
OK. Not sure what the pit refers to in the case of the virtual system.
The rtc can only be read in one second chunks.
This does not matter, some radio clocks allow the same.
The only disadvantage is that when the step time back must
be done on the Host, the /dev/rtc gets stuck (it is a monotonic
clock) and the qemu must be restarted.
rtc is not a monotonic clock. It can be set to whatever time you
want.
Unless your hardware is different than what I am imagining.
This comes from the implementation of the /dev/rtc by the
qemu. I haven't investigated why.
Regards
Pavel
Regards
Pavel
-----Original Message-----
From: unruh [mailto:un...@wormhole.physics.ubc.ca]
Sent: Tuesday, June 15, 2010 2:23 AM
To: questions@lists.ntp.org
Subject: Re: [ntp:questions] Reference clock driver for /dev/rtc
On 2010-06-14, Krejci, Pavel
<pavel.kre...@siemens-enterprise.com> wrote:
Hello,
I have written the reference clock driver for /dev/rtc on
Linux. We use it to synchronize the guest Linux system
running in the
qemu with the Host clock. If this is useful to someone
else I would
like to contribute to the NTP project.
How should I proceed?
Why would you want to do that? The rtc is almost certainly
worse than
the system clock. Why not have the guest just read the
host's system
clock, rather than the rtc.
_______________________________________________
questions mailing list
questions@lists.ntp.org
http://lists.ntp.org/listinfo/questions
_______________________________________________
questions mailing list
questions@lists.ntp.org
http://lists.ntp.org/listinfo/questions
_______________________________________________
questions mailing list
questions@lists.ntp.org
http://lists.ntp.org/listinfo/questions
