GPS disciplined DCXO with software temperature compensation and IEEE 1588 PTP master
Already have an account? Log in.
To make the experience fit your profile, pick a username and tell us what interests you.
Redesign of GPSDO Master Clock
Key Features:
The MAX6577 temperature sensor appears to exhibit both Gaussian and burst noise. The Gaussian noise is desirable since it improves the temperature measurement precision with appropriate filtering, but the burst noise is slightly problematic. The burst noise produces abrupt 10 mC to 20 mC steps in the temperature measurement which result in 2 ppb to 5 ppb perturbations of the frequency correction. These perturbations produce bursts of minor offset drift.
I've replaced the internal crystal oscillator and temperature sensor with an external oscillator and analog temperature sensor. This provides both better precision for temperature measurements as well as tighter thermal coupling between the oscillator and the temperature sensor.
The frequency output of the MAX6577 provides better than 0.01 C precision when combined with 32-bit timer capture. This allows for a PPS RMS tracking error of less than 10 ns.
I've successfully synchronized an embedded PTP client with the GPS time server. I still need to verify that I'm correctly setting some of the protocol fields, but I've otherwise managed to get +/- 100ns tracking over a single switch hop using the IEEE 802.1AS gPTP profile. The gPTP profile uses raw ethernet frame encapsulation with the 0x88F7 ether type. I decided to use this particular profile since it provides a simple zeroconf style configuration.
On the client side, I've implemented the offset tracking as 1 Hz update process with three stages:
The result is a surprisingly clean and consistent offset tracking error.
Robert Rouquette
These sudden shifts need to be mitigated to reduce tracking error in NTP and PTP clients.
MAC Address Sync Delay
54:55:58:D2:83:2F -3 8.264 us
offset measurement:
- used: 6
- drift: 0.083 ppm
- mean: 0.029 us
- dev: 0.032 us
drift measurement:
- used: 16
- mean: -0.005 ppm
- dev: 0.100 ppm
I've rewritten the frequency and offset correction code to be entirely digital and restructured the system clock to be a three layer structure with a free-running monotonic timescale, a frequency disciplined timescale, and an offset disciplined timescale. The three-layer clock allows for independent disciplining of frequency and timing offset and improves overall tracking accuracy. The system is now capable of frequency steering with a resolution of 0.232 ppb.
In addition, I've restructured the GPSDO and NTP code into a single unit to make GPS/NTP fail-over more seamless and improve code reuse. NTP source selection and health tests are still a WIP, but I've also implement support for select chronyc queries:
This allows for viewing fairly detailed status information with the chronyc utilty.
A large benefit of this massive rewrite is that the frequency error measurement is now both precise and accurate enough to expose the presence of hysteresis loops in the tracking error. Further analysis is needed to isolate the cause, but they still make some interesting patterns. The loop widths are less than 30ppb, and they were previously buried in the measurement noise.
After exhaustively comparing the datasheets and performing software tests on different eval boards, it's become clear to me that these two parts are functionally identical down to the pinout, register map, and address map. The only difference I've been able to find is the onboard ROM.
Thankfully, since I'm not using the ROM, I can use these parts interchangeably which makes sourcing them significantly easier.
Updates:
Human readable status examples:
$ echo -n 'ntp' | nc -uW1 -w1 #.#.#.# 23
gps epoch: 0xE7423F96.71E6AA43
ntp offset: 0xFFFFFFEE.00000000
ntp servers: pool.ntp.org
address stratum leap reach next offset (ms) delay (ms) jitter (ms) drift (ppm) skew (ppm) weight
- 38.229.60.9 2 0 77777 8 2.750 29.478 0.328 -0.142 6.376 0.1159
- 159.89.86.140 2 0 77777 5 7.713 23.209 0.390 0.317 8.137 0.0000
- 144.172.118.20 2 0 77777 10 0.260 9.997 0.190 0.267 4.714 0.1568
- 45.79.214.107 2 0 77777 7 2.613 7.500 0.404 1.649 8.263 0.0895
- 162.159.200.1 3 0 77777 4 0.576 10.040 0.249 0.512 4.612 0.1603
- 23.131.160.7 2 0 77777 10 5.772 29.849 0.178 0.402 3.595 0.2056
- 74.6.168.73 2 0 77777 7 3.304 35.071 0.310 0.232 6.270 0.1179
- 155.248.196.28 2 0 66767 4 0.970 25.825 0.194 -0.397 4.354 0.1540
clock stratum: 1
clock offset: 2.411 ms
clock drift: 0.304 ppm
leap indicator: 0
$ echo -n 'gpsdo' | nc -uW1 -w1 #.#.#.# 23 pps locked: yes current offset: -88 ns mean offset: -17.6 ns rms offset: 42.9 ns frequency skew: 0.0215 ppm frequency correction: -175.9287 ppm - pll trimming: 0.0839 ppm compensation: -176.0126 ppm - temperature: 37.834 C - bias: 37.230 C - coefficient: 0.4296 ppm/C - offset: -176.2722 ppm
Temperature Compensation Settling:
I've implemented SNMP and the GPSDO logic. The software is currently using the frequency trimming support in the PTP hardware to under-clock the PPS and steer the edge alignment. Otherwise, the temperature compensation and PLL logic are working as expected. I still need to make a breakout board for the DXCO and I2C temperature sensor.
Temperature Compensation:
I've decided to modify the design to use the TM4C1294 series MCU. It has essentially the same feature as the MSP432, but with a higher clock speed and it is currently in-stock at multiple suppliers.
The CPU clock speed is nominally rated for 120MHz, but I have so far seen no issues overclocking it to 125 MHz to get a clean 8ns of timer resolution. The 32-bit timers also seem to have no issue operating at 125 MHz.
I've added an e-ink display to the design for debug and status purposes, and also implemented DHCP and NTP.
PPS Tracking:
Comparison against GPS NTP server:
I've been using an EK-TM4C1294XL eval board for prototyping. I'm currently working on an eval board layout for the DCXO and temperature sensor. The GPS eval board should be arriving sometime next week.