[ntp:questions] NTP.ORG MEINBERG KEEP TIME ACCURATE to 10MS
unruh at invalid.ca
Thu Nov 15 08:23:14 UTC 2012
On 2012-11-15, David Taylor <david-taylor at blueyonder.co.uk.invalid> wrote:
> On 15/11/2012 07:22, unruh wrote:
>> Actually more like 2 microseconds accuracy should be possible, unless
>> the Raspberry Pi's servicing of the GPIO interrupt is really very bad.
>> Nice page.
> Glad you liked the write-up. Just some notes for myself, really!
> If you search for "latency" here:
> "jbeale" says: "My experiments with NTP-PPS timing in the 3.1.9 kernel
> with a relatively unloaded system indicate that my R-Pi has an interrupt
> latency ranging between 10 and 80 microseconds, and occasionally gets as
> long as 150 microseconds (see below)"
Hm. Looks pretty bad.
I once did tests with a 300MHz Linux machine, where I would send out a
signal on one of the output pins on the parallel port, reading the time
at which I did so, and then read in the time on the Ack interrupt driven pin
and found delays typically of 1-2 microseconds (timing only to the
microsecond due to the clock reading function I used). That he seems to
be betting up to 150us delay is really bad.
> and provides a graph of some values he has seen. I haven't made any
> direct measurements. On my system here, the offsets reported by NTP are
> less, but it would not show any fixed interrupt latency. It's all well
> within the 10 milliseconds requirement, of course.
> with glitches due to restarts. Zero offset is plotted as 20 microseconds.
More information about the questions