[ntp:questions] Is 24PPM an Excessive Real-Time Clock Correction?
root at localhost.invalid
Mon Jul 9 14:07:52 UTC 2007
Richard B. Gilbert wrote:
> David T. Ashley wrote:
>> I have ntpd running on a RHEL Linux Dell 1U rack server in another city.
>> I have no experience with other servers (to know how accurate the
>> clocks are or are not).
>> The value of "drift" that gets settled on after a few days is 24 PPM
>> (about 14 seconds/week).
>> This strikes me as a little high, because even my $20 Timex watch does
>> better than this.
>> Any insight into whether 24 PPM is excessive for my server?
> Consider that a computer manufacturer typically spends something like $2
> US on the components for the clock!
You say "the clock" like as if there were only one.
I count at least 6 "clocks" on a typical x86-based platform:
o PIT (Programmable Interval Timer) e.g. Intel's 8253 and 8254
o RTC (Real-Time Clock)
o (Local??) APIC timer (not too sure about this one)
o ACPI timer, also known as PM (Power Management) clock
o CPU TSC (timestamp counter)
o HPET (High Precision Event Timers)
I might have missed some?
As far as I understand, some manufacturers are pushing hard to
have high-quality HPET integrated into every south bridge.
The minimum requirements for the HPET are rather underwhelming:
Fmin = 10 MHz
Clock Frequency Drift
+- .05 % (500 ppm) Over any interval >= 1 ms
+- .2 % (2000 ppm) Over any interval <= 100 µs
But I've heard some manufacturers (IBM?) provide high-quality HPETs.
(Could anyone comment on this aspect?)
More information about the questions