ntpd with Intel SpeedStep and AMD PowerNow, any updates? - NTP

This is a discussion on ntpd with Intel SpeedStep and AMD PowerNow, any updates? - NTP ; Given that almost all new desktop, laptop, and server hardware (AMD, Intel, even the new UltraSPARC T1) is now shipped with some form of variable clock-speed power management enabled, it would seem that ntpd needs some method of coping with ...

+ Reply to Thread
Results 1 to 6 of 6

Thread: ntpd with Intel SpeedStep and AMD PowerNow, any updates?

  1. ntpd with Intel SpeedStep and AMD PowerNow, any updates?

    Given that almost all new desktop, laptop, and server hardware (AMD,
    Intel, even the new UltraSPARC T1) is now shipped with some form of
    variable clock-speed power management enabled, it would seem that ntpd
    needs some method of coping with this feature sensibly.

    Looking at
    http://ntp.isc.org/bin/view/Support/...#Section_9.1.2.
    and also the list archives, it seems there is no current solution
    other than to disable such features or run ntp periodically in a cron
    job.

    Given that these features can cycle the processor clock speedf by a
    factor of two or more, isn't this still a big problem? A look at the
    changelog included with 4.2.5p19 didn't seem to indicate that any
    fixes for this were in place.

    Wouldn't falling back to using the (limited resolution) of the timer
    interrupt in the presence of SpeedStep or PowerNow be good default
    behavior?

    All of my current servers with -like features run w32time on Windows,
    but we are going to be introducing some new Linux servers soon, and I
    am concerned that I will be effectively without ntpd on these boxes.

    Regards,
    Ryan


  2. Re: ntpd with Intel SpeedStep and AMD PowerNow, any updates?

    >>> In article <1174433788.522437.176240@d57g2000hsg.googlegroups. com>, "Ryan Malayter" writes:

    Ryan> Given that almost all new desktop, laptop, and server hardware (AMD,
    Ryan> Intel, even the new UltraSPARC T1) is now shipped with some form of
    Ryan> variable clock-speed power management enabled, it would seem that ntpd
    Ryan> needs some method of coping with this feature sensibly.

    I'm thinking this is more of an OS issue than an ntpd issue.

    The timekeeping APIs are pretty well defined, and ntp relies on the OS being
    able to "keep" time.

    But that's my opinion.

    What do you suggest?

    H

  3. Re: ntpd with Intel SpeedStep and AMD PowerNow, any updates?

    On Tue, 20 Mar 2007, Ryan Malayter wrote:
    > Given that almost all new desktop, laptop, and server hardware (AMD,
    > Intel, even the new UltraSPARC T1) is now shipped with some form of
    > variable clock-speed power management enabled, it would seem that ntpd
    > needs some method of coping with this feature sensibly.
    >
    > Looking at
    > http://ntp.isc.org/bin/view/Support/...#Section_9.1.2.
    > and also the list archives, it seems there is no current solution
    > other than to disable such features or run ntp periodically in a cron
    > job.


    Hmm, I'm not seeing any issues with AMD PowerNow and ntpd on my Linux box.
    I use that feature and my ntpd works just fine.

    $ uname -a
    Linux jaipur 2.6.20-gentoo-r1 #1 Wed Mar 7 12:03:52 PST 2007 x86_64 AMD
    Athlon(tm) 64 Processor 3500+ AuthenticAMD GNU/Linux

    $ ntpq -p
    jaipur:~> ntpq -p
    remote refid st t when poll reach delay offset
    jitter
    ================================================== ============================
    *cpe-66-91-253-2 .GPS. 1 u 712 1024 377 13.629 1.245
    1.142
    -nodem-rsm-vl99. 132.239.1.6 2 u 774 1024 377 18.432 3.488
    0.305
    +wcsdca1-gsr1.sa 66.75.160.254 2 u 782 1024 377 8.005 -0.004
    1.450
    -seattle.esysmai 129.6.15.29 2 u 756 1024 377 44.151 -3.309
    1.140
    -n118.ewd.goldma 128.194.254.9 3 u 666 1024 377 52.720 -4.973
    3.100
    +gatekeeper.no-s .TRUE. 1 u 150 1024 377 49.808 -1.101
    33.099

    $ ntptime
    ntp_gettime() returns code 0 (OK)
    time c9ab5581.d1ee4000 Tue, Mar 20 2007 23:56:01.820, (.820042),
    maximum error 948348 us, estimated error 1170 us
    ntp_adjtime() returns code 0 (OK)
    modes 0x0 (),
    offset 2352.000 us, frequency -31.776 ppm, interval 1 s,
    maximum error 948348 us, estimated error 1170 us,
    status 0x1 (PLL),
    time constant 10, precision 1.000 us, tolerance 512 ppm,

    $ ntpq -c rv
    assID=0 status=06f4 leap_none, sync_ntp, 15 events, event_peer/strat_chg,
    version="ntpd 4.2.4@1.1437-o Sun Mar 4 00:20:25 UTC 2007 (1)",
    processor="x86_64", system="Linux/2.6.20-gentoo-r1", leap=00, stratum=2,
    precision=-20, rootdelay=13.629, rootdispersion=48.268, peer=13255,
    refid=66.91.253.27,
    reftime=c9ab4e9a.51ab6ceb Tue, Mar 20 2007 23:26:34.319, poll=10,
    clock=c9ab55df.304075ca Tue, Mar 20 2007 23:57:35.188, state=4,
    offset=2.620, frequency=-31.776, jitter=1.746, noise=1.170,
    stability=0.232, tai=0

    $ stats
    Current: 1.00
    Freq Ticks
    1.00 108218232 ( 95.5%)
    1.80 114953 ( 0.1%)
    2.00 12276 ( 0.0%)
    2.20 4963612 ( 4.4%)

    This last script shows the number of kernel ticks the processor has spent
    at each clock speed, in GHz. I use the built-in Linux "ondemand"
    processor scaling governor.

    -Ryan



  4. Re: ntpd with Intel SpeedStep and AMD PowerNow, any updates?

    On Mar 21, 2:01 am, Ryan Moore wrote:
    > Hmm, I'm not seeing any issues with AMD PowerNow and ntpd on my Linux box.
    > I use that feature and my ntpd works just fine.


    I did some testing with ntpd 4.2.0a and Ubuntu 6.10 (kernel 2.6.17.11)
    on my Pentium-M laptop (with 'ondemand' processor scaling enabled),
    and everything seemed to work well, even when I preiodically stopped
    and started the cpuburn utility to force the processor to scale
    speeds. ntpd did not panic and exit, and jitter was fairly low.

    So it would appear we have at least two data points that indicate
    frequency scaling is no longer an issue for ntpd, at least with more
    recent Linux kernels.

    I suspect, as Harlan suggested, that it was an OS issue with older
    kernels using the time stamp counter instead of the PM timer in the
    kernel time APIs.

    Do you think I should update the wiki with this information, or should
    I wait for further tests on other hardware platforms and Linux distros
    (we're planning on RHEL for the new servers)?

    Regards,
    Ryan


  5. Re: ntpd with Intel SpeedStep and AMD PowerNow, any updates?

    On 2007-03-23, Ryan Malayter wrote:

    > I suspect, as Harlan suggested, that it was an OS issue with older
    > kernels using the time stamp counter instead of the PM timer in the
    > kernel time APIs.
    >
    > Do you think I should update the wiki with this information, or should
    > I wait for further tests on other hardware platforms and Linux distros
    > (we're planning on RHEL for the new servers)?


    Go head and add the information that you have. Just be sure to clearly
    identify the NTP and Linux kernel versions, and hardware, that this
    information is applicable to.

    --
    Steve Kostecke
    NTP Public Services Project - http://ntp.isc.org/

  6. Re: ntpd with Intel SpeedStep and AMD PowerNow, any updates?

    On Mar 23, 2:09 pm, Steve Kostecke wrote:
    > Go head and add the information that you have. Just be sure to clearly
    > identify the NTP and Linux kernel versions, and hardware, that this
    > information is applicable to.


    Done. Mr. Moore, would you like to add your working configuration as
    well?

    Also, if I wrote any incorrect assumptions into the Wiki (for example,
    the kernel is using the PM timer, which I read about on an AMD tech
    support page), please let me know and I will fix or remove them.

    Thank you,
    Ryan


+ Reply to Thread