Re: [Linuxptp-users] Improved accuracy with software timestamps under high CPU load

[Miroslav L. <mli...@re...>]

On Wed, Jul 23, 2014 at 02:36:24AM +0000, Dylan Davis wrote:
> I was experimenting to see how far I could push PTP accuracy using software
> timstamps. I have a simple test setup with a single master and a single slave,
> configured to use the P2P delay mechanism. logSyncInterval on the master is set
> to -8.

-8 is an extremenly short update interval, I'm a bit surprised it still
works. Did you use the default PI configuration?

> Both systems have tickless kernel disabled (nohz=off). With both systems
> idle I can reach around 2us rms offset and 15-20us peak offset. If I then fully
> load all 4 cores of the slave (using the 'stress' program), after a brief spike
> the accuracy of the synchronization actually *improves*, reaching 300-500ns rms
> offset and around 3us peak offset. Removing the CPU load brings the offset back
> to its original magnitude. In both cases the rms and max offset values are
> consistent, with only occasional spikes. Interestingly, the reported delay under
> load is 12us less than when unloaded.

Beside what Richard has mentioned, I think it could be also the
variability in the time it takes the CPU to wake up from various
powersaving states. You can try setting the idle= or
intel_idle.max_cstate= kernel options and see if that helps.

-- 
Miroslav Lichvar