rt_clk_sync

Hard Real-Time world-wide clock synchronization utilizing new better algorithms than PTP and NTP uses.

Objectives

• "NTP without PLL" -- continue the ideas of [1], [2]

  • better stability (no feedback with obscured interoperation
    between digital quasi-VCO inside of OS which nobody
    understands, and the ad-hoc tuned PI-controller within ntpd in
    userspace
  • possibly better performance
    • provide means to tune system to particular process (computer
      clock) & measurement (network) noises by means of their
      absolute variances as well as temporal characteristics
    • according to control theory laws, optimally tuned estimator
      will perform same or better than any possible feedback system
      for the NTP task

• establish better NTP-like userspace program <-> OS interface

  • provide more clear i/f based on current time offset & drift
    estimates input into OS's timekeeping core;
    no P, I controller parameters, no ad-hoc quasi-VCOs;
    the Linux' adjtimex() is an example of devil's evil
  • try Linux as the reference implementation
  • join efforts with PTPd guys, they probably need exactly the
    same

• allow simultaneous locking to several NTP servers

  • AFAIK impossible with current ntpd client, correct me if wrong

• implement optimal linear estimator for clock processes, allowing to run with no absolute master at all [3], [4], [5]

In longer term:

• adaptive estimation of both clock and network parameters (autotuning, or continuous adaptation)

• revise/redesign application interface to allow e.g. delay/latency measurement between two hosts (us and some other) on demand or reliable reporting of current/min/max accuracy (delay/latency) which all other connected stations have

• statistics/logging of estimated accuracy (for audits, revisions, etc.)

The ultimate goal (megalomanic):

• release the Internet-wide network of computers and their clocks,
  managed as a non-hierarhical cloud of clocks of very different
  stochastical properties (from GPSes and highly stable atomic
  standards to ordinary PC's crystals);
  apply clock ensembling algorithms to perform this large-scale
  time fusion

To Do

Outline

Test & prototype the idea in two different scenarios:

• LAN, 1 NTP server
• WAN, multiple NTP servers

Tasks

• write clk_sync_observer, preferably in Linux' user-space

  • should run besides classical, non-patched ntpd
  • should monitor bidirectional NTP network datagrams, parse them
    and record their timestamps wrt. non-disciplined local clock
  • record evolution of disciplined ("ntp") time wrt. local clock

• provide means to perform/record simultaneous NTP queries to serveral NTP servers from fixed set

• define reference PC or "PC-like" HW platform

  • Ethernet card
  • some reasonable, but not unrealistically good local clock
    (i.e. no undeterministic, unbeatable errors, but on the other
    hand no unfairly good oscillators, common consumer/industrial
    HW only)
  • a means to output OS' clock synchronous marks, at least
    (at worst) using GPIO

• make a simple and portable FPGA & Rb clock reference measurement kit (for this purpose, 10ns is fairly good, so no fancy interpolation techniques from VZLU's TDC are needed)

• perform measurements

  • WAN scenario
  • LAN scenario

• evaluate acquired data

  • perform offline simulations of proposed algorithms
  • compare against bare NTP

• patch Linux' timekeeping.c

  • wipe out PI/quasi-VCO, replace with offset&drift estimates
    • for compatibility reason, current PI/VCO mess may perhaps
      become an upper shell on top of the offset/drift i/f

• write a NTP client-side userspace replacement, using algorithms and new OS i/f as described

  • first, only single-server mode
  • later, add multiple-servers fusion
  • (much later: adaptive filtering...)

Resources

• https://github.com/jeremyfix/easykf (C++ Kalman filtering (Extended Kalman Filter, Unscented Kalman Filter)
• https://github.com/TKJElectronics/KalmanFilter (Kalman filter used to calculate the angle, rate and bias from from the input of an accelerometer/magnetometer and a gyroscope)

People & Credits

• Dave Mills (for NTP)
• Thomas Gleixner (for current timekeeping.c)
• Pavel Pisa (guidance to tglx's timekeeping.c)
• Jan Pacner (ptpd port to QNX, started ntpd vs. ptpd comparisons)
• Marek Peca (noPLL NTP idefix, clock ensembling algorithms)

[1] J.Levine: Synchronizing computer clocks using Kalman filters. 43rd PTTI Meeting, 2011.

[2] J.Ridoux, D.Veitch: Ten Microseconds Over LAN, for Free. IEEE ISPCS, 2007.

[3] K.R.Brown: The Theory of the GPS Composite Clock. Proceedings of ION GPS-91, 1991.

[4] C.A.Greenhall: A Review of Reduced Kalman Filters for Clock Ensembles. IEEE UFFC 2012.

[5] M.Peca, V.Michalek, M.Vacek: Clock Composition by Wiener Filtering Illustrated on Two Atomic Clocks. EFTF 2013 (accepted)