Skip to content

README.md

Latest commit

 

History

History
147 lines (98 loc) · 9.21 KB

README.md

File metadata and controls

147 lines (98 loc) · 9.21 KB
Raw

Chrony NTP Setup for Strom

Install and configure chrony on a strom host with broadcast/video-production-friendly settings: tight polling, multi-source corroboration, and the TAI/UTC leap table installed so userspace reading CLOCK_TAI gets the correct offset (~37 s vs UTC, not 0).

Why chrony, not systemd-timesyncd

Ubuntu/Debian default to systemd-timesyncd, which has two problems for any host running broadcast-style timestamping:

  1. Single source. timesyncd polls one NTP server at a time. With no corroborating peers, the kernel's maxerror field grows linearly between polls — on a host with the default fallback (ntp.ubuntu.com) and a max poll of ~34 min, maxerror reaches the ~1 s range routinely. Strom's System Clock panel reads this directly and reports Health=Degraded as a result.
  2. No TAI/UTC leap table. Without leapsectz right/UTC installed, the kernel's TAI−UTC offset stays at 0. Code reading CLOCK_TAI (PTP-aware paths, audio/video timestamping) is silently 37 s off. UTC-only monitoring will never surface this.

chrony fixes both: multi-source voting (with minsources 3), proper outlier rejection, and leapsectz right/UTC to populate the TAI table.

The broader rationale follows Ateliere Live's recommendation for video production: https://help.ateliere.com/live/docs/installation/base-platform/ntp/. The Ubuntu NTP pool deliberately does not apply leap smearing, which is a feature for CLOCK_TAI consumers — a smeared second stretched across many hours produces a non-integer TAI/UTC offset and userspace drifts away from reality for the duration of the smear. A clean 1 s step is preferable.

Install

sudo bash scripts/setup/ntp/install-chrony.sh

The script:

  1. Installs the chrony package (which auto-masks systemd-timesyncd on Ubuntu/Debian, and pulls in tzdata-legacy on Ubuntu 24.04+ for the right/UTC zoneinfo).
  2. Backs up any existing /etc/chrony/chrony.conf to /etc/chrony/chrony.conf.pre-ateliere (won't overwrite an existing backup on re-run).
  3. Writes the broadcast/video-production config (see below) to /etc/chrony/chrony.conf.
  4. Restarts the chrony service.
  5. Prints chronyc tracking + chronyc sources -v after an 8 s warmup.

Idempotent — safe to re-run.

What the config does

The full conf is embedded in the script. Key choices beyond Ateliere's baseline:

  • Four Ubuntu NTP pools (ntp.ubuntu.com, 0/1/2.ubuntu.pool.ntp.org) with iburst. Provides chrony with enough sources to satisfy minsources 3 and to vote out falsetickers.
  • minpoll 6 / maxpoll 8 on every pool — pins the steady-state poll to 64 s..256 s. Default maxpoll 10 (1024 s) is unnecessarily loose for broadcast. Public NTP pool guidance asks for minpoll 6 or slower, which is honored here.
  • minsources 3 — do not update the clock unless at least three sources agree. Blocks single-source false updates.
  • leapsectz right/UTC — installs the TAI/UTC offset table so userspace reading CLOCK_TAI gets the correct offset. Required for any pipeline doing TAI/PTP-aware timestamping.
  • log measurements statistics tracking selection — per-sample and per-selection entries in /var/log/chrony/ so a falseticker is attributable after the fact.

Verify

From the strom UI

Open the Clocks panel. After the switch, the System Clock should show:

  • Health: Healthy (was Degraded with timesyncd)
  • TAI − UTC offset: 37 s (was 0 s — the silent footgun)
  • Max error: low ms range (was hundreds of ms with timesyncd)

From the host

chronyc tracking
chronyc sources -v
chronyc sourcestats

chronyc sources -v should show one ^* (selected best), one or more ^+ (combined), and the rest ^- / ^?. No x (falseticker) or ~ (too variable).

Production deployment results (2026-05-05)

Replaced systemd-timesyncd with this config on a strom production host (Ubuntu 24.04.3 LTS, kernel 6.8, hosted environment with sub-10 ms RTT to several stratum-2 servers).

strom GUI System Clock panel — before vs. after:

Field timesyncd (before) chrony (after)
Health Degraded Healthy
TAI − UTC offset 0 s 37 s
Current offset -11.97 µs 0 µs
Estimated error 0 µs ¹ 519 µs
Max error 940000 µs 4785 µs

¹ timesyncd does not populate the kernel's estimated-error field meaningfully — it stays at 0.

Two practically important changes:

  1. TAI − UTC went 0 → 37 s. The leapsectz right/UTC payoff. Code paths reading CLOCK_TAI were previously reporting times 37 s wrong with no warning, because UTC-only monitoring cannot surface this. Single most important reason to run chrony with leapsectz on any host doing broadcast-style timestamping.
  2. Max error dropped ~200× (940 ms → 4.8 ms). This is what flipped Health to Healthy. timesyncd couldn't shrink max error because it had only one source; chrony with four pools and minsources 3 produces a real bounded error from corroborating samples.

PLL active: no after the switch is not a regression — chrony disciplines the system clock via adjtimex frequency steering rather than enabling the kernel PLL loop. By design and generally considered more accurate.

chronyc tracking ~8 s after restart showed 77 µs RMS already — better than the home-LAN baseline below at T ≈ 5 min, reflecting the better network position.

Reference baseline (home-LAN + fiber + public internet)

For comparison, a one-off baseline test of the same config on a consumer-grade setup — useful to gauge what's achievable without local stratum-1 infrastructure.

Test environment

  • Debian 12 (bookworm), kernel 6.1, consumer x86_64 workstation
  • Home LAN behind a consumer fiber line, over public internet to the Ubuntu NTP pool — no local stratum-1, no PTP
  • Chrony 4.3 from Debian bookworm

Results

Three measurements after a clean systemctl restart chrony:

Metric T ≈ 5 min T ≈ 20 min T ≈ 37 min
RMS offset 427 µs 228 µs 167 µs
Last offset 68 µs 235 ns 30.6 µs
System time vs. reference 56 ns 1.5 µs 13.9 µs
Skew 0.94 ppm 0.165 ppm 0.092 ppm
Residual freq 0.028 ppm 0.000 ppm -0.004 ppm
Update interval 65 s 130 s 261 s
Root dispersion 229 µs 305 µs 459 µs

Sources at T ≈ 37 min:

  • 8/8 sources reachable (reach 377 octal = last 8 polls all successful)
  • 1 source selected as best (^*), a Nordic stratum-2 server with ~65 µs per-sample std dev
  • 7 sources marked not-combined (^-) — all measured within 0.5–3 ms of the selected source, but with std dev ranging from ~70 µs up to ~1.5 ms
  • No falsetickers (x), no too-variable (~)
  • Polls on stable sources reached log2 = 8 (256 s) = the configured maxpoll. The log2 = 6 (64 s) minpoll was only active during the first few minutes of convergence

Conclusions from the baseline

  1. Practical sync ceiling over home LAN + fiber + public internet is ≈100–200 µs RMS offset. The floor is set by WAN jitter to the best single source (~65 µs std dev). No amount of chrony tuning will beat that floor without local time infrastructure.
  2. Frequency stability is effectively at the limit of a consumer PC crystal: skew 0.092 ppm after ~37 min. Kernel clock drift between polls is negligible; chrony is not the bottleneck.
  3. Chrony does not combine sources (^+) in this setup — the best source is ~15× better than the next best, so combining would degrade, not improve, accuracy. This is correct behavior. Seeing ^+ requires multiple sources of comparable quality, which over public internet typically means multiple LAN-reachable servers.
  4. Outlier visibility is cheap. Adding log selection to /var/log/chrony/selection.log costs essentially nothing and turns a "why did the clock jump" post-mortem from guesswork into a lookup.

Going lower than 100 µs

The only path to sub-100 µs over this kind of network is local time infrastructure:

  • GPS-disciplined stratum-1 on the LAN: a dedicated box with a GPS/GNSS receiver acting as NTP stratum 1. Puts the jitter floor around 10–50 µs depending on switch quality.
  • PTP (IEEE 1588) with hardware timestamping across PTP-aware switches: sub-microsecond is realistic. Requires NIC and switch support end-to-end.
  • For a video production context, PTP is the standard (SMPTE 2059-2 profile). NTP is the baseline; PTP is the target for frame-accurate lock.

Rollback

sudo apt-get purge -y chrony
sudo systemctl enable --now systemd-timesyncd

The pre-Ateliere chrony.conf (if any existed) is preserved at /etc/chrony/chrony.conf.pre-ateliere.

References