[ntp:questions] questions] Performance estimation

Miroslav Lichvar mlichvar at redhat.com
Wed Jun 17 13:03:12 UTC 2020


On Wed, Jun 17, 2020 at 09:09:22AM +0100, David J Taylor wrote:
> I don't know what the path is on the RPi for the PPS interrupt but I would
> expect it to be similar to that on the BBblack, so that latency graph is of
> interest - thanks!

FWIW, there is a polling version of the GPIO PPS driver, which avoids
the interrupt latency with a busy loop and which might be useful for
measuring the latency. It will make the system clock more accurate,
but it might have a negative impact on the accuracy of the NTP server
as the PPS interrupt latency compensates the NIC interrupt latency
(when the NIC driver is not polling). In my tests with a RPi 3B+ the
RX timestamping error was about 20us larger than the TX error.

The driver is here: https://github.com/mlichvar/pps-gpio-poll

Getting it to work with the Raspbian kernel might be tricky. IIRC I
didn't work for me the last time I tried. I prefer OpenWrt on these
small computers.

-- 
Miroslav Lichvar



More information about the questions mailing list