Kernel clock issue "Clocksource tsc unstable"

[NoOp]

On 02/27/2009 07:46 PM, Rashkae wrote:
> NoOp wrote:
>> On 02/27/2009 06:55 PM, Rashkae wrote:
>>> NoOp wrote:
>>>
>>>> I recall that thread:
>>>> http://comments.gmane.org/gmane.linux.ubuntu.user/145397
>>>> [Computer loosing time]
>>>> might be worth looking at again.
>>>>
>>>>
>>> Not really,, That thread led me to believe the problem was with my power
>>> supply, which I replaced, and declared success.  However, my success was
>>> very short lived.  It was afterwards that I discovered I can fix it by
>>> changing my clocksource from tsc to acpm, but this was something so
>>> obscure, (and it seemed, not bothering anyone else) that I never
>>> bothered posting about it.
>>>
>> 
>> But shouldn't it switch automatically to acpi_pm if tsc is not stable?[1]
>> 
>> For instance, on the Thinkpad A21M that I receive the dmesg of:
>> Clocksource tsc unstable (delta = -87685142 ns)
> 
> yes, it should.  but stable doesn't really have much to do with
> accurate... TSC is not accurate, it's not even meant to be.  (I'm not
> sure exactly what kinds of problems make it unstable.  I used to think
> it was synchronization between cpu cores, but apparently, there is more
> to it than that.)
> 
> On systems where it works, however, TSC is very performance friendly.  I
> guess the reasoning is, having an 8 core super server with dozens of
> vm's is all fine and good, unless all your processes get stuck
> contending for the one timer chip on your mobo.
> 
> 

Found out what TSC stands for: Time Stamp Counter

http://www.google.com/linux?hl=en&q=tsc&btnG=Search

<http://www.redhat.com/docs/en-US/Red_Hat_Enterprise_MRG/1.1/html/Realtime_Tuning_Guide/sect-Realtime_Tuning_Guide-Realtime_Specific_Tuning-TSC_timer_synchronization_on_Opteron_CPUs.html>

<quote>
3.4. TSC timer synchronization on Opteron CPUs
The current generation of AMD64 Opteron processors are susceptible to a
large gettimeofday skew when cpufreq is enabled while using the Time
Stamp Counter (TSC). MRG Realtime provides a method to prevent this on
Opteron systems by forcing all processors to simultaneously change to
the same frequency. As a result, the TSC on a single processor never
increments at a different rate than the TSC on another processor.
</quote>

$ man gettimeofday

is interesting... again, don't understand it all, but it is interesting :-)