Re: [Linuxptp-users] clock_gettime performance using ptp hw clock

[Richard C. <ric...@gm...>]

On Thu, Jul 25, 2013 at 11:36:22AM +0000, Ledda William EXT wrote:

> Surprisingly I have found a quite high result: the mean value to
> read the PTP clock is about 5 us. To read the time CLOCK_REALTIME or
> CLOCK_MONOTONIC clocks, I obtain a mean value of about 600 ns!!

(It is not a surprise to me ;)

The i210 is on a PCIe bus, which is a serial bus with a packet based
protocol. Individual operations (think PEEK and POKE) have a high
latency due to protocol overhead. Five microseconds sounds about
right.

The Linux kernel offers a highly optimized path to the system
clocks. On some architectures it even maps kernel time keeping
variables directly to a user space page (vdso) in order to allow low
latency time stamps.

> Can somebody had experience about this? Is there some other way to
> read the PTP time with sub-microseconds accuracy?

I wrote a whole paper about this. If you want sub-microsecond
accuracy, then you want to use using hardware signals, not
software. The best real time systems I have seen (Xenomai or
PREEMPT_RT) have a worst case latency of about 10 microseconds.

   http://ieeexplore.ieee.org/xpl/articleDetails.jsp?arnumber=6070158

The best way I know is to synchronize the Linux system time to the PTP
hardware clock using the phc2sys program. Applications can then just
use CLOCK_REALTIME or CLOCK_MONOTONIC without any changes to the code.

HTH,
Richard