Skip to content

Research data: time transfer, laboratories and Internet measurements

ntpstats reads the data behind recent time-transfer and NTP measurement studies directly:

Format Where it comes from What you get
cggtts GNSS time-transfer receivers at time laboratories (BIPM CGGTTS V2E) REF − GNSS time per epoch; ntpstats cv gives UTC(A) − UTC(B)
rinex-clock IGS clock products (.clk, gzip accepted) receiver and satellite clock biases
circular-t BIPM Circular T, section 1 UTC − UTC(k) for about 80 laboratories
ripe-atlas RIPE Atlas NTP measurements (API JSON) offset and RTT per probe and server
ntppool NTP Pool monitoring logs offset, RTT and score per monitor
interop the ntpstats open dataset (data/interop/, monthly since 2026) NTP/NTS/NTPv5/Roughtime probes of public servers, with availability

All are auto-detected and work with every command: info, stability, noise, hat and report.

Stability of UTC(k) from Circular T

Circular T gives UTC − UTC(k) every five days. Concatenate the issues you have into one file:

curl -O https://webtai.bipm.org/ftp/pub/tai/Circular-T/cirt/cirt.44[0-9]
cat cirt.4* > cirt-all.txt
ntpstats info cirt-all.txt --all-peers                   # one series per laboratory
ntpstats stability cirt-all.txt --peer PTB -k tdev,htot  # stability of UTC(PTB)

The published uncertainties (uA, uB, u) are kept in each series' metadata.

GNSS common view between two laboratories

ntpstats cv laba.cggtts labb.cggtts -o ab.csv           # REF(A) - REF(B), common view
ntpstats cv laba.cggtts labb.cggtts --mode aiv -o ab.csv # all in view
ntpstats stability ab.csv -k tdev,mdev
  • Common view differences the same satellite at the same epoch, so satellite clock errors cancel.
  • All in view differences the per-epoch means.
  • Line checksums of each file are verified; failing lines are skipped and counted.

GNSS timing receivers (u-blox UBX)

A raw u-blox receiver log (binary UBX frames, possibly mixed with NMEA sentences) is read directly (format ubx, detected automatically). Three messages are used, with the layouts of the u-blox interface description:

Message Series and columns
UBX-NAV-CLOCK receiver clock bias as reported (clkB), with clock_drift, time_accuracy (tAcc), freq_accuracy (fAcc)
UBX-NAV-TIMEUTC the UTC time of each navigation epoch (time stamps), utc_accuracy
UBX-TIM-TP the quantization error qErr of each time pulse, at the pulse time

Without NAV-TIMEUTC, times come from TIM-TP's GPS week with GPS − UTC = 18 s.

Removing the PPS sawtooth

The receiver can only start its PPS on an edge of its internal clock, so each pulse is off by up to half a clock period, drifting in a sawtooth; qErr predicts it. ntpstats sawtooth pairs each time-interval-counter reading of the PPS with the qErr of that pulse and removes it:

ntpstats sawtooth examples/data/pps-tic.csv examples/data/ubx-timing.ubx -o corrected.csv
# 1800 PPS samples paired with qErr, sign +1 (data (smaller sample-to-sample variance))
#    TDEV  τ=1s 2.51 ns -> 1.16 ns  τ=10s 592 ps -> 378 ps  ...

u-blox does not state the sign with which qErr applies; ntpstats tries both and keeps the one that removes the sawtooth (the smaller sample-to-sample variance), and reports it. Fix it with --sign +1 or --sign -1 once it is known for a receiver and counter. The sawtooth notebook walks through the example.

gpsd: the PPS as the host sees it

Most Linux GNSS time servers run gpsd, which feeds chrony or ntpd. Its JSON stream reports every PPS edge twice: as GNSS time (real_sec, real_nsec) and as the system clock time-stamped it (clock_sec, clock_nsec). Recorded with gpspipe -w, or sampled live:

gpspipe -w > gpspipe.json                       # leave it running, then
ntpstats stability gpspipe.json --peer PPS -k tdev,mtie
ntpstats watch gpsd -i 1 -o gpsd-pps.csv        # live, from localhost:2947 (--command-override host:port)
ntpstats sawtooth gpspipe.json --peer PPS       # receivers that report qErr: no UBX log needed

Each device gives a PPS series (GNSS minus the system clock's time stamp of the edge, the ntpstats sign convention) and a TOFF series (the same for the serial message, with its much larger latency), with precision and, when the receiver reports it, qerr.

TAPR TICC

The TAPR TICC timestamping counter prints 1.897999794440 chA lines (Timestamp, Period and 3-Corner-Hat modes) or bare numbers (Time Interval mode). Each channel becomes the phase of its input against an ideal grid of its period (its time interval error, measured by the TICC's reference); with two channels, chA - chB of the events paired within half a period is added. The seconds printed with the WRAP setting are unwrapped, and seconds and fractions are kept apart so that the picoseconds survive. Times are counter seconds since the TICC started.

ntpstats stability ticc-log.txt --peer "chA - chB" -k adev,tdev
ntpstats info interval-log.txt -f ticc          # Time Interval mode: bare numbers need -f ticc

Clock products (IGS)

ntpstats stability IGS0OPSFIN_20250190000_01D_05M_CLK.CLK.gz --peer PTBB -k oadev,hdev

NTP servers seen from many vantage points (RIPE Atlas)

curl "https://atlas.ripe.net/api/v2/measurements/MSM_ID/results/?format=json&start=...&stop=..." -o atlas.json
ntpstats info atlas.json --all-peers

In Python, ntpstats.research.group_summary(series, by="peer") gives the offset distribution per server (median, 95th percentile of |offset|) across all probes, or by="probe" per vantage point. Atlas logs local − server; ntpstats negates it to its usual server − local.

NTP Pool monitors

curl "https://www.ntppool.org/scores/192.0.2.1/log?limit=2000&monitor=*" -o pool.csv
ntpstats info pool.csv --all-peers        # one series per monitor
ntpstats hat pool.csv --all-peers         # which monitor paths are noisy?

Please keep requests to public services modest.

The ntpstats open interop dataset

Every week the Live interop workflow probes public NTP, NTS, NTS-pool, NTPv5 and Roughtime servers; the first run of each month is recorded in data/interop/. The dataset is versioned with every release (Zenodo), and its schema is in the directory's README.

ntpstats dataset data/interop --test nts      # availability and offsets of NTS servers
ntpstats dataset data/interop --trend         # each server over the runs
ntpstats stability data/interop --peer time.google.com -k oadev

A directory given to any command is read as the concatenation of its files.

Synthetic examples of every format are in examples/data/.