Building off my last NTP post (Microsecond accurate NTP with a Raspberry Pi and PPS GPS), which required a $50-60 GPS device and a Raspberry Pi (also $40+), I have successfully tested something much cheaper, that is good enough, especially for initial PPS synchronization. Good enough, in this case, is defined as +/- 10 milliseconds.
How accurate of time do you really need? The last post showed how to get all devices on a local area network (LAN) within 0.1 milliseconds of “real” time. Do you need you equipment to be that accurate to official atomic clock time (12:03:05.0001)? Didn’t think so. Do you care if every device is on the correct second compared to official/accurate time (12:03:05)? That’s a lot more reasonable. Using a u-blox USB GPS can get you to 0.01 seconds of official. The best part about this? The required USB GPS units are almost always less than $15 and you don’t need a Raspberry Pi.
This post will show how to add a u-blox USB GPS module to NTP as a driver or chrony (timekeeping daemon) as a reference clock (using GPSd shared memory for both) and verify the accuracy is within +/- 10 milliseconds.
USB u-blox GPS (VK-172 or GT-U7), GT-U7 preferred because it has a micro-USB plug to connect to your computer. It is important to note that both of these are u-blox modules, which has a binary data format as well as a high default baudrate (57600). These two properties allow for quick transmission of each GPS message from GPS to computer.
1 – Update your host machine and install packages
This tutorial is for Linux. I use Ubuntu so we utilize Aptitude (apt) for package management:
In /etc/default/gpsd, change the settings to the following:
# Start the gpsd daemon automatically at boot time
# Use USB hotplugging to add new USB devices automatically to the daemon
# Devices gpsd should collect to at boot time.
# this could also be /dev/ttyUSB0, it is ACM0 on raspberry pi
# -n means start listening to GPS data without a specific listener
Reboot with sudo reboot.
3a – Chrony configuration (if using NTPd, go to 3b)
I took the default configuration, added my 10.98 servers, and more importantly, added a reference clock (refclock). Link to chrony documentation here. Arguments/parameters of this configuration file:
iburst means send a bunch of synchronization packets upon service start so accurate time can be determined much faster (usually a couple seconds)
maxpoll (and minpoll, which isn’t used in this config) is how many seconds to wait between polls, defined by 2^x where x is the number in the config. maxpoll 6 means don’t wait more than 2^6=64 seconds between polls
refclock is reference clock, and is the USB GPS source we are adding
‘SHM 0’ means shared memory reference 0, which means it is checking with GPSd using shared memory to see what time the GPS is reporting
‘refid NMEA’ means name this reference ‘NMEA’
‘offset 0.000’ means don’t offset this clock source at all. We will change this later
‘precision 1e-3’ means this reference is only accurate to 1e-3 (0.001) seconds, or 1 millisecond
‘poll 3’ means poll this reference every 2^3 = 8 seconds
‘noselect’ means don’t actually use this clock as a source. We will be measuring the delta to other known times to set the offset and make the source selectable.
[email protected]:~ $ sudo cat /etc/chrony/chrony.conf
# Welcome to the chrony configuration file. See chrony.conf(5) for more
# information about usuable directives.
#pool 2.debian.pool.ntp.org iburst
server 10.98.1.1 iburst maxpoll 6
server 10.98.1.198 iburst maxpoll 6
server 10.98.1.15 iburst
refclock SHM 0 refid NMEA offset 0.000 precision 1e-3 poll 3 noselect
Restart chrony with sudo systemctl restart chrony.
3b – NTP config
Similar to the chrony config, we need to add a reference clock (called a driver in NTP). For NTP, drivers are “servers” that start with an address of 127.127. The next two octets tell what kind of driver it is. The .28 driver is the shared memory driver, same theory as for chrony. For a full list of drivers, see the official NTP docs. To break down the server:
‘server 127.127.28.0’ means use the .28 (SHM) driver
minpoll 4 maxpoll 4 means poll every 2^4=16 seconds
noselect means don’t use this for time. Similar to chrony, we will be measuring the offset to determine this value.
‘fudge 127.127.28.0’ means we are going to change some properties of the listed driver
‘time1 0.000’is the time offset calibration factor, in seconds
‘stratum 2’ means list this source as a stratum 2 source (has to do with how close the source is to “true” time), listing it as 2 means other, higher stratum sources will be selected before this one will (assuming equal time quality)
Running gpsmon shows us generally information about the GPS, including time offset. The output looks like the below screenshot. Of importance is the satellite count (on right, more is better, >5 is good enough for time), HDOP (horizontal dilution of precision) is a measure of how well the satellites can determine your position (lower is better, <2 works for basically all navigation purposes), and TOFF (time offset).
In this screenshot the TOFF is 0.081862027, which is 81.8 milliseconds off the host computer’s time. Watch this for a bit – it should hover pretty close to a certain value +/- 10ms. In my case, I’ve noticed that if there are 10 or less satellites locked on, it is around 77ms. If there are 11 or more, it is around 91ms (presumably due to more satellite information that needs to be transmitted).
5 – record statistics for a data-driven offset
If you are looking for a better offset value to put in the configuration file, we can turn on logging from either chrony or NTPd to record source information.
Edit /etc/chrony/chrony.conf and uncomment the line for which kinds of logging to turn on:
# Uncomment the following line to turn logging on.
log tracking measurements statistics
Then restart chrony (sudo systemctl restart chrony) and logs will start writing to /var/log/chrony (this location is defined a couple lines below the log line in chrony.conf):
[email protected]:~ $ ls /var/log/chrony
measurements.log statistics.log tracking.log
For NTPd (be sure to restart it after making any configuration changes):
[email protected] ~ % cat /etc/ntp.conf
# Enable this if you want statistics to be logged.
statistics loopstats peerstats clockstats
filegen loopstats file loopstats type day enable
filegen peerstats file peerstats type day enable
filegen clockstats file clockstats type day enable
Wait a few minutes for some data to record (chrony synchronizes pretty quick compared to NTPd) and check the statistics file, filtered to our NMEA refid:
cat /var/log/chrony/statistics.log | grep NMEA
This spits out the lines that have NMEA present (the ones of interest for our USB GPS). To include the headers to show what each column is we can run
cat /var/log/chrony/statistics.log | head -2; cat /var/log/chrony/statistics.log | grep NMEA
# ntp, there is no header info so we can omit that part of the command
cat /var/log/peerstats | grep 127.127.28.0
NTP stats don’t include header information. The column of interest is the one after the 9014 column. The columns are day, seconds past midnight, source, something, estimated offset, something, something, something. We can see the offset for this VK-172 USB GPS is somewhere around 76-77 milliseconds (0.076-0.077 seconds), which we can put in place of the 0.000 for the .28 driver for NTP and remove noselect.
So now we have some data showing the statistics of our NMEA USB GPS NTP source. We can copy and paste this into Excel, run data to columns, and graph the result and/or get the average to set the offset.
This graph is certainly suspicious (sine wave pattern and such) and if I wasn’t writing this blog post, I’d let data collect overnight to determine an offset. Since time is always of the essence, I will just take the average of the ‘est offset’ column (E), which is 7.64E-2, or 0.0763 seconds. Let’s pop this into the chrony.conf file and remove noselect:
Restart chrony again for the config file changes to take effect – sudo systemctl restart chrony.
6 – watch ‘chrony sources’ or ‘ntpq -pn’ to see if the USB GPS gets selected as the main time source
If you aren’t aware, Ubuntu/Debian/most Linux includes a utility to rerun a command every x seconds called watch. We can use this to watch chrony to see how it is interpreting each time source every 1 second:
# for chrony
watch -n 1 chronyc sources
In the above screenshot, we can see that chrony actually has the NMEA source selected as the primary source (denoted with the *). It has the Raspberry Pi PPS NTP GPS ready to takeover as the new primary (denoted with the +). All of the sources match quite closely (from +4749us to – 505us is around 5.2 milliseconds). The source “offset” is in the square brackets ([ and ]).
# for ntp
watch -n 1 ntpq -pn
7- is +/- five millseconds good enough?
For 99% of use cases, yes. You can stop here and your home network will be plenty accurate. If you want additional accuracy, you are in luck. This GPS module also outputs a PPS (pulse per second) signal! We can use this to get within 0.05 millseconds (0.00005 seconds) from official/atomic clock time.
In this post, we got a u-blox USB GPS set up and added it as a reference clock (refclock) to chrony and demonstrated it is clearly within 10 millisecond of official GPS time.
You could write a script to do all this for you! I should probably try this myself…
In the next post, we can add PPS signals from the GPS module to increase our time accuracy by 1000x (into the microsecond range).
A note on why having faster message transmission is better for timing
My current PPS NTP server uses chrony with NMEA messages transmitted over serial and the PPS signal fed into a GPIO pin. GPSd as a rule does minimum configuration of GPS devices. It typically defaults to 9600 baud for serial devices. A typical GPS message looks like this:
$GPGGA, 161229.487, 3723.2475, N, 12158.3416, W, 1, 07, 1.0, 9.0, M, , , , 0000*18
That message is 83 bytes long. At 9600 baud (9600 bits per second), that message takes 69.1 milliseconds to transmit. Each character/byte takes 0.833 milliseconds to transmit. That means that as the message length varies, the jitter will increase. GPS messages do vary in length, sometimes significantly, depending on what is being sent (i.e. the satellite information, $GPGSV sentences, is only transmitted every 5-10 seconds).
I opened gpsmon to get a sample of sentences – I did not notice this until now but it shows how many bytes each sentence is at the front of the sentence:
These sentences range from 83 bytes to 35 bytes, a variation of (83 bytes -35 bytes)*0.833 milliseconds per byte = 39.984 milliseconds.
Compare to the u-blox binary UBX messages which seem to always be 60 bytes and transmitted at 57600 baud, which is 8.33 milliseconds to transmit the entire message.
The variance (jitter) is thus much lower and can be much more accurate as a NTP source. GPSd has no problem leaving u-blox modules at 57600 baud. This is why the USB GPS modules perform much more accurate for timekeeping than NMEA-based devices when using GPSd.
For basically every GPS module/chipset, it is possible to send it commands to enable/disable sentences (as well as increase the serial baud rate). In an ideal world for timekeeping, GPSd would disable every sentence except for time ($GPZDA), and bump up the baud rate to the highest supported level (115200, 230400, etc.). Unfortunately for us, GPSd’s default behavior is to just work with every GPS, which essentially means no configuring the GPS device.
Lots of acronyms in that title. If I expand them out, it says – “microsecond accurate network time protocol with a Raspberry Pi and pulse per second global positioning system [receiver]”. What it means is you can get super accurate timekeeping (1 microsecond = 0.000001 seconds) with a Raspberry Pi and a GPS receiver that spits out pulses every second. By following this guide, you will your very own Stratum 1 NTP server at home!
Why would you need time this accurate at home?
You don’t. There aren’t many applications for this level of timekeeping in general, and even fewer at home. But this blog is called Austin’s Nerdy Things so here we are. Using standard, default internet NTP these days will get your computers to within a 10-20 milliseconds of actual time (1 millisecond = 0.001 seconds). By default, Windows computers get time from time.windows.com. MacOS computers get time from time.apple.com. Linux devices get time from [entity].pool.ntp.org, like debian.pool.ntp.org. PPS gets you to the next SI prefix in terms of accuracy (milli -> micro), which means 1000x more accurate timekeeping.
You should update your Pi to latest before basically any project. We will install some other packages as well. pps-tools help us check that the Pi is receiving PPS signals from the GPS module. We also need GPSd for the GPS decoding of both time and position. I use chrony instead of NTPd because it seems to sync faster than NTPd in most instances and also handles PPS without compiling from source (the default Raspbian NTP doesn’t do PPS) Installing chrony will remove ntpd.
In /boot/config.txt, add ‘dtoverlay=pps-gpio,gpiopin=18’ to a new line. This is necessary for PPS. If you want to get the NMEA data from the serial line, you must also enable UART and set the initial baud rate.
sudo bash -c "echo '# the next 3 lines are for GPS PPS signals' >> /boot/config.txt"
sudo bash -c "echo 'dtoverlay=pps-gpio,gpiopin=18' >> /boot/config.txt"
sudo bash -c "echo 'enable_uart=1' >> /boot/config.txt"
sudo bash -c "echo 'init_uart_baud=9600' >> /boot/config.txt"
In /etc/modules, add ‘pps-gpio’ to a new line.
sudo bash -c "echo 'pps-gpio' >> /etc/modules"
2 – wire up the GPS module to the Pi
I used the Adafruit Ultimate GPS breakout. It has 9 pins but we are only interested in 5. There is also the Adafruit GPS hat which fits right on the Pi but that seems expensive for what it does (but it is significantly neater in terms of wiring).
GPS PPS to RPi pin 12 (GPIO 18)
GPS VIN to RPi pin 2 or 4
GPS GND to RPi pin 6
GPS RX to RPi pin 8
GPS TX to RPi pin 10
see 2nd picture for a visual
Now place your GPS antenna (if you have one) in a spot with a good view of the sky. If you don’t have an antenna, you might have to get a bit creative with how you locate your Raspberry Pi with attached GPS.
I honestly have my antenna in the basement (with only the kitchen and attic above) and I generally have 8-10 satellites locked all the time (11 as of writing). Guess that means the antenna works better than expected! Either way, better exposure to the sky will in theory work better. Pic:
Edit /etc/default/gpsd and change GPSD_OPTIONS=”” to GPSD_OPTIONS=”-n” and change DEVICES=”” to DEVICES=”/dev/ttyS0 /dev/pps0″, then reboot. My full /etc/default/gpsd is below:
[email protected]:~ $ sudo cat /etc/default/gpsd
# Default settings for the gpsd init script and the hotplug wrapper.
# Start the gpsd daemon automatically at boot time
# Use USB hotplugging to add new USB devices automatically to the daemon
# Devices gpsd should collect to at boot time.
# They need to be read/writeable, either by user gpsd or the group dialout.
# Other options you want to pass to gpsd
4.5 – check GPS for good measure
To ensure your GPS has a valid position, you can run gpsmon or cgps to check satellites and such. This check also ensures GPSd is functioning as expected. If your GPS doesn’t have a position solution, you won’t get a good time signal. If GPSd isn’t working, you won’t get any updates on the screen. The top portion will show the analyzed GPS data and the bottom portion will scroll by with the raw GPS sentences from the GPS module.
5 – edit config files
For chrony, add these two lines to the /etc/chrony/chrony.conf file somewhere near the rest of the server lines:
My entire /etc/chrony/chrony.conf file looks like this:
###### below this line are custom config changes #######
server 10.98.1.1 iburst minpoll 3 maxpoll 5
server time-a-b.nist.gov iburst
# delay determined experimentally by setting noselect then monitoring for a few hours
# 0.325 means the NMEA time sentence arrives 325 milliseconds after the PPS pulse
# the delay adjusts it forward
refclock SHM 0 delay 0.325 refid NMEA
refclock PPS /dev/pps0 refid PPS
allow 10.98.1.0/24 # my home network
###### above this line are custom config changes #######
###### below this line is standard chrony stuff #######
#log tracking measurements statistics
makestep 1 3
Restart chrony, wait a few minutes, and verify.
sudo systemctl restart chrony
5 – verify the NTP server is using the GPS PPS
Right after a chrony restart, the sources will look like this (shown by running ‘chronyc sources’)
For the S column, * means the source that is active. + means it is considered a good source and would be used if the current one is determined to be bad or is unavailable. The x shown for the NMEA source means it is a “false ticker”, which means it isn’t being used. In our case that is fine because the PPS source is active and valid. Anything else generally means it won’t be used.
In this case, chrony is using the PPS signal. The value inside the brackets is how far off chrony is from the source. It is showing that we are 289 nanoseconds off of GPS PPS time. This is very, very close to atomic clock level accuracy. The last column (after the +/-) includes latency to the NTP source as well as how far out of sync chrony thinks it is (for example, the 220.127.116.11 server is 12.5 milliseconds away one-way via ping):
[email protected]:~ $ ping -c 5 18.104.22.168
PING 22.214.171.124 (126.96.36.199) 56(84) bytes of data.
64 bytes from 188.8.131.52: icmp_seq=1 ttl=54 time=25.2 ms
64 bytes from 184.108.40.206: icmp_seq=2 ttl=54 time=27.7 ms
64 bytes from 220.127.116.11: icmp_seq=3 ttl=54 time=23.8 ms
64 bytes from 18.104.22.168: icmp_seq=4 ttl=54 time=24.4 ms
64 bytes from 22.214.171.124: icmp_seq=5 ttl=54 time=23.4 ms
--- 126.96.36.199 ping statistics ---
5 packets transmitted, 5 received, 0% packet loss, time 4007ms
rtt min/avg/max/mdev = 23.403/24.954/27.780/1.547 ms
A day after the bulk of writing this post, I turned on logging via the .conf file and restarted chrony. It settled to microsecond level accuracy in 57 seconds. For the offset column, the scientific notation e-7 means the results are in nanoseconds. This means that for the 19:54:35 line, the clock is -450.9 nanoseconds off of the PPS signal. There is that e-10 line in there which says that it is 831 picoseconds off PPS (I had to look up SI prefixes to make sure pico came after nano! also I doubt the Pi can actually keep track of time that closely.). After the initial sync, there is only 1 line in the below log that is actually at the microsecond level (the others are all better than microsecond) – the 20:00:59 entry, which shows the clock is -1.183 microseconds off.
Things that affect timekeeping on the Pi
Thought I’d toss in this section for completeness (i.e. thanks for all the good info Austin but how can I make this even better?). There are a few things that affect how well the time is kept on the Pi:
Ambient temperature around the Pi – if you plot the freq PPM against ambient temperature, there will be a clear trend. The more stable the ambient temp, the less variation in timekeeping.
Load on the Pi – similar to above. Highly variable loads will make the processor work harder and easier. Harder working processor means more heat. More heat means more variability. These crystals are physical devices after all.
GPS reception – they actually making timing GPS chips that prefer satellites directly overhead. They have better ability to filter out multipathing and such. In general, the better the GPS reception, the better the PPS signal.
After running through the steps in this guide, you should now have a functional Stratum 1 NTP server running on your Raspberry Pi with microsecond level accuracy provided by a PPS GPS. This system can obtain time in the absence of any external sources other than the GPS (for example, internet time servers), and then sync up with the extremely precise GPS PPS signal. Our NTP GPS PPS clock is now within a few microseconds of actual GPS time.