[time-nuts] Re: PPS latency? User vs kernel mode

[folkert]

> > I wrote something this evening that just busy-loops waiting for a raising
> > edge and then writes down the current time to feed that to ntp. 
> 
> You are trading interrupt latency for user-kernel crossing delays and 
> scheduler quirks.

Are you sure about the user-kernel crossings? Because (not 100% sure) I
thought wiringpi access the gpio pins directly - iirc they're memory
mapped (strace doesn't show any system calls) on ARM processors.

> There is probably a way to dedicate a CPU to your polling job.  I think I've 

https://unix.stackexchange.com/questions/530575/how-to-dedicate-a-cpu-to-a-process/530579
Will try that.

> What ntpd are you using?  The PPS driver may have an option to log every 
> pulse.  It's FLAG4 in ntp classic and ntpsec.  That would let you compare your 
> two paths.

I'm using ntpsec. Unfortunately I don't have any hardware (yet) for
comparing pulses so I depend on ntp telling me if it is a false-ticker
or not.

> You also need to compare the adev type statistics from the interrupt path when 
> your polling is/isn't running to see if polling is distorting the interrupt 
> path.

I understand the neccessity for the adev calculation. I don't understand
the remark about the interrupt distortion: if I use the poller-program,
then I don't enable (via the pps-gpio module) interrupts?