Hi Patrick,<br><br>I only just had a chance to glance at that paper. Thanks for the link (and the reference therein to my e-mail on the topic).<br><br>I think the results are somewhat dependent on the time scale. From your table 2, system-load-full gave standard deviations on the order of ~ms. In that case I am sure the time-stamping mechanism is less important, which is confirmed indirectly by your system-load-cpu figure being 1000x better. The network prioritization and unfiltered spikes of MtSD is dominant.<br>
<br>Did you get to try any experiments to improve the convergence rate ? Just last week I decided to tweak the time interval in timer.c (in the 'old ptpd')<br><br> itimer.it_value.tv_sec = itimer.it_interval.tv_sec = 0;<br>
itimer.it_value.tv_usec = itimer.it_interval.tv_usec = 100000;<br><br>And found my nodes claim to converge in < 45 sec. I have yet to do any real test (with external pulses applied to measure the true time offsets) but it seems a simple solution to getting my system ready faster.<br>
<br>Cheers<br><br>Gertjan<br><br><br><br><div class="gmail_quote">On Sun, Jan 31, 2010 at 11:49 PM, Patrick Ohly <span dir="ltr"><<a href="mailto:patrick.ohly@intel.com">patrick.ohly@intel.com</a>></span> wrote:<br>
<blockquote class="gmail_quote" style="border-left: 1px solid rgb(204, 204, 204); margin: 0pt 0pt 0pt 0.8ex; padding-left: 1ex;"><div class="im">On Fr, 2010-01-29 at 17:15 +0000, gertjan hofman wrote:<br>
> We didn't change any of the time stamping mechanism but found that we<br>
> could get improvement by taking a closer look at the filtering applied<br>
> to the Master to Slave Delay etc. Especially with store and forward<br>
> switches, it is not obvious to me that the kernel time stamping is<br>
> the dominating factor in the time synchronization.<br>
<br>
</div>Hardware time stamping on the host definitely led to a measurable<br>
improvement:<br>
<a href="http://www.linuxclustersinstitute.org/conferences/archive/2008/PDF/Ohly_92221.pdf" target="_blank">http://www.linuxclustersinstitute.org/conferences/archive/2008/PDF/Ohly_92221.pdf</a><br>
<br>
But I agree, delays inside the network are more likely to affect the<br>
results, which is why PTP v2 support in ptpd and v2 enabled switches are<br>
important. I'm sure your traffic prioritization also helps.<br>
<font color="#888888"><br>
--<br>
Best Regards, Patrick Ohly<br>
<br>
The content of this message is my personal opinion only and although<br>
I am an employee of Intel, the statements I make here in no way<br>
represent Intel's position on the issue, nor am I authorized to speak<br>
on behalf of Intel on this matter.<br>
<br>
<br>
</font></blockquote></div><br><br clear="all"><br>-- <br>==================================================<br>Gertjan Hofman<br>ghofman [at] <a href="http://gmail.com">gmail.com</a> gertjan.hofman [at] <a href="http://honeywell.com">honeywell.com</a><br>
<br>604-982-3574<br>==================================================<br>