Internal Clock Drift Estimation in Computer Clusters

Hal Murray hmurray at
Thu Aug 25 21:50:19 UTC 2016

>> That's from 2008.
>> Was there anything in there that you thought was particularly interesting 
>> surprising?

> The actual numbers.  I've seen the logic before.

That was a long time ago.  Any numbers are likely to be misleading.  Best to 
collect your own.  How about "Trust, but verify."

PC have two crystals, one for the CPU and one for the RTC/TOY clock.  In the 
old days of that paper, Linux used an interrupt from the RTC clock for 
timekeeping.  The TSC was used to interpolate between ticks but the main 
timekeeping use the RTC crystal.  Current Linux timekeeping uses the TSC 
which is derived from the CPU crystal.

The RTC crystal probably has a much better temp coefficient than the CPU 
crystal and/or it is probably farther from the CPU so gets less temperature 
change under load.

There is another worm in this can.  On most PCs, the CPU clock is "spread" to 
dance under the FCC EMI rules.  I don't have any data on the temperature 
coefficient of that stuff.  (SoC boards like the Raspberry Pi typically don't 
do that.)

They used 100 megabit ethernet.  Anything more advanced than a Raspberry Pi 
has gigabit these days.

These are my opinions.  I hate spam.

More information about the devel mailing list