[LinuxPPS] So far, so good

[Hal V. Engel]

On Monday 09 August 2010 12:52:30 pm Miroslav Lichvar wrote:
> > My system once it has settled down will typically 
> > show offsets in the +-250 nano seconds range when I query it with
> > ntptime.  If  I increase the system load (by running a large build for
> > example) the offsets will increase and then settle down again once ntp
> > has adjusted for the higher latency of the interrupt handler.
> 
> Isn't that a change in frequency caused by increased temperature
> rather than increased latency?
> 
> A shorter polling interval might help in such situations, but it will
> probably make things worse when idle.
> 


Both probably come into play.  If it were mostly temperature it would take a 
little while for the oscillator to be affected but I find that my system will 
start showing increased offsets with in a few seconds of the change in system 
load and then stabilize in a few minutes.  If the primary issue was 
temperature related then I would expect it to take several minutes for the 
temperature of the oscillator to change enough to cause this but since it is 
almost instantaneous I think it is mostly caused by the increased latency of 
the interrupt handler.   Also if I do a single threaded build on my dual core 
system I don't see the quick change in the offset that I do with a multi-
threaded build.  Of course the single threaded build generates less heat but 
it also will have much less impact on the latency of the interrupt handler 
since one core is sitting idle and can service the interrupt very quickly.

Of course we can do little more than speculate on this since we don't have any 
way to measure the interrupt handler latency.  About the only way to test this 
would be to use a machine with a temperature controlled oscillator as a test 
bed.  If loading such a machine causes significant changes to the offset then 
that affect can probably be attributed to latency rather than temperatures.  
But I don't have a machine like this to test with.

Hal