[CentOS-virt] Status of kernel divider option in 2.6.9-78.0.1

Wed Sep 17 18:56:56 UTC 2008
Jeff <jlar310 at gmail.com>

On Wed, Sep 17, 2008 at 1:35 PM, Akemi Yagi <amyagi at gmail.com> wrote:
> On Wed, Sep 17, 2008 at 11:23 AM, Jeff <jlar310 at gmail.com> wrote:
>> Under 4.6, we recompiled the kernel with HZ=100 for improved
>> time-keeping in VMware guests. I've read about the backporting of the
>> divider patch into RHEL/CentOS 4.7, but it sounds like it also comes
>> with some bugs. I have been unable to determine the current status of
>> the divider option in the latest 4.7 kernel update. I have
>> experimented with "divider=10" and it works with no major problems,
>> except that the guest clock drifts ahead very slightly. So far, it
>> looks like several seconds of drift per day. Not a show-stopper, but
>> still a concern. With HZ=100, our time sync was spot on. In both cases
>> we are using "clock=pit nosmp noapic" and tools.syncTime="TRUE". If it
>> matters, our host machines are CentOS 4.6 x86_64.
>>
>> Can anyone comment on the current status of bugs with the divider
>> option? Is anyone getting accurate time sync with divider=10 with
>> 2.6.9-78.0.1? Is there anything else to watch out for? Is there any
>> fix for the slight time drift with divider=10?
>>
>> I am primarily concerned with i686.
>>
>> Currently, the SRPM for -78.0.1 is missing from the mirrors, so if I
>> want to update, I have to rely on the divider option for my VMware
>> guests. It also appears that Tru has not yet added any -78 RPMs to his
>> kernel-vm project.
>
> As far as I can tell from my limited experience, the clock issue
> occurs regardless of the method taken, namely, kernel-vm (100Hz
> kernel) or divider=10. (Of course, this is with older kernels, not
> -78)  Both options work fine when it comes to idle %cpu.  One
> improvement of the 4.7 kernel is that it does not crash even when
> divider= is used together with clock=pit like it does with CentOS-5
> kernels.
>
> Yes, the srpm file for -78.0.1 is missing (see
> http://bugs.centos.org/view.php?id=3132 ).  But we can try -78 for
> performing a test.

Our experience with HZ=100 in 4.6 kernels was nearly perfect time
sync, at least to within a second over one week. Not so with
divider=10. I will grab the -78 SRPM and give it a spin.

-- 
Jeff