Precision Time Protocol daemon

Hello all,

after looking at PTPd as possible replacement for NTP in HPC-style clusters ("Beowulf") we decided to continue that discussion with a broader audience of people building such clusters. PTPd looked useful (thanks a lot for creating it!), but it needs to be more visible to find its way into more installations. I mailed the Beowulf mailing list [1] today; hopefully that will lead to a lively discussion.

Kendall, it would be great to have you or anyone else reading this forum and interested in HPC join the discussion. I also pointed out that developing patches for PTPd and submitting them for inclusion might be worthwhile (for example for the next revision of IEEE 1588); I hope that's okay.

[1] http://dir.gmane.org/gmane.comp.clustering.beowulf.general

It seems like your post generated some interest. I just got back to working on PTPd, so I'm sorry I didn't participate. One quick comment. I saw a post about mentioning data acquisition as a good application for 1588. That and control systems were the original motivation for 1588. Since control systems usually run on RTOSes, that probably wouldn't be relevant to HPC. Correlating distributed acquisition data, however, seems very relevant. Keep us updated on your progress, and thanks for your interest.

I completely forgot to keep you informed about was going on with "PTP for HPC" in the last year. Time to change that, because quite a few things have happened!

First, Intel has released a mainstream Ethernet NIC (82576, http://sourceforge.net/project/showfiles.php?group_id=42302&package_id=283702), which supports hardware time stamping. As an Intel employee I was able to get access to early hardware samples. Then I modified drivers and PTPd to use this new hardware capability. In contrast to the MPC8313E I decided against an ifdef-based approach and instead moved the hardware access into a platform specific layer. I also wanted to experiment with system time synchronization, so I implemented different modes of operation (synchronizing NIC time vs. synchronizing system time with two different approaches). A full report about that work was presented at the LCI 2008 conference: http://www.linuxclustersinstitute.org/conferences/archive/2008/PDF/Ohly_92221.pdf

Second, Intel is about to approve publishing the patches. They contain some minor bug fixes (IMHO), the code restructuring and the Intel specific code. The hope is that these patches can be incorporated into PTPd trunk but obviously I'll have to publish them for review before we can discuss that.

Third, there was a discussion on the Linux netdev list about a generic kernel interface for hardware time stamping of network packets. Search for Octavian Purdila: he championed that idea and wants to work on the kernel side of things. The intention at Intel is to support that development by adapting the Intel driver and PTPd to the new kernel interface.

All in all the future for PTP in HPC looks good. Getting all components into the hands of actual users is still a challenge, but the pieces are starting to come together.

Bye, Patrick

Patrick,

Any idea when the 82576 will be released to the public?

Thanks,
Matt

Matt, I believe it has been released. On the other hand I don't know where such hardware is normally sold. I can try to find out, if you want to.

In related news, I got permission to publish the PTPd patches. I'll upload them next Monday.

I posted the patches in the tracker. For those who prefer to get a complete copy of the sources with the patches included: they are now also in github.

http://github.com/pohly/ptpd/commits/hw-2008-11-07-lci-2008