NIST NTP servers
mel at beckman.org
Wed May 11 14:27:47 UTC 2016
Most data centers will require a remotely positioned NTP server, which is actually easier and cheaper than a remotely located active GPS antenna. I have placed the $300 commercial NTP servers in an environmental box on the roof, powering t by PoE, without problems.
You don't need a redundant network either, nor redundant power. Just plunk down two of these gizmos, or as I suggested elsewhere, deploy one or more CDMA, GSM, or WWV-based clocks, for as much local infrastructure and signal source diversity as you like (I sourced some of these units earlier in the thread, all well less than $1K each. You pay more for diversity, but you get more too.
In response to the several DIYers on this thread: Anyone who thinks they're building a raspberry pi-based GPS NTP server for just $150 is kidding themselves. They forgot to include their labor, which is worth far more than the $300 for a commercial unit. The same goes for people who complain about even the minimal $300 price, forgetting that a commercial product has to pay for marketing, support, and make a profit.
External NTP has two kinds of vulnerabilities: the ones we know, and the ones we don't. The ones we know are serious enough the pat the ones we don't should be considered with respect. Maybe diversity in Internet sources is a cure, maybe it isn't. But diverse RF sources is demonstrably more secure than the Internet.
My point stands: secure external RF NTP sources are so plentiful that relying on Internet NTP is just plain crazy.
> On May 11, 2016, at 7:12 AM, Andreas Ott <andreas at naund.org> wrote:
>> Boss: That sounds expensive. How much are we talking?
>> IT guy: $300
> Over the past year we made engineering samples to deploy to datacenters.
> The goal was to use GPS and PPS to discipline ntpd appliances and serve
> as stratum 1 to other NTP distribution servers without the $5k price tag
> of commercial NTP 1RU gear. We also deliberately not pursued the path of
> running antenna coax from the roof to a receiver, as that is not an
> option in all the datacenters where we need to deploy.
> These appliances were built for lesss than $150 as
> (a) Raspberry-Pi with GPS receiver board
> (b) Garmin 18(x) LVC with DB-9 to an "older whitebox server"
> In my experience, most locations inside datacenters where you have good
> power and network connectivity have not "good enough" GPS signal reception
> due to servers emitting lots of RF noise in the range 1-2 GHz on the
> L-band. A brand new suite in the datacenter had OK GPS quality, but
> once we added 20+ racks with 40 servers each, GPS reception was pretty
> much gone. This hardware was an active antenna with less than 6 feet of
> cabling routed to the top of the network cabling rack. Most smartphones
> can run an app to show you the GPS signal on the phone, just walk around
> your datacenter and compare the signal.
> The only workable solution was to move the GPS clock to a location
> where it had good GPS signal but neither redundant network nor conditioned
> power (aka. "on my desk near a south facing window"). It also works pretty
> well "in my garage".
> In places where GPS reception is good, you can achieve 10E-06 seconds
> accuracy over time even with cheap hardware. If you chose to run the DB-9
> NMEA0183 and GPS as "serial port passthrough" to a VM on a Hypervisor
> you can still get better than 10E-03 seconds accuracy.
> Andreas Ott (Time-Nut) K6OTT +1.408.431.8727 andreas at naund.org
More information about the NANOG