METHOD, SYSTEM AND DEVICE FOR CLOCK SYNCHRONIZATION OVER TIME-VARYING AND LOSSY NETWORKS

Abstract

The present invention is directed towards a method, a system, and a device which allow an improved synchronization of a system-wide timing information over a network. Hence, slave clocks can be synchronized to a high quality clock, such as a master clock.

Description

The present invention is directed towards a method which allows a synchronization of client clocks providing the master time to several client clocks. A server machine runs a reference clock and forms the basis for a system-wide time, which therefore has to be distributed over a network. The present invention provides an improved mechanism which allows a coupling of several client machines to the specific server machine with high synchronization accuracy. The present invention is also directed towards a system providing a client clock and a server clock, which are synchronized. Furthermore, a client is suggested which can be run in several instances distributed over a network, which is then synchronized with this server time.

For many applications, clocks at distributed locations must be synchronized. This can be done by exchanging time messages over a communication network connecting the clocks. However, time-varying behaviour of the communication network affects the synchronization performance. This invention concerns methods to mitigate the effects of unreliable communication networks on synchronization performance.

One of the known methods is the Network Time Protocol NTP, which is a networking protocol for synchronizing clocks among computer systems over packet-switched, variable-latency data networks. The basic idea of NTP is to provide time stamps for sending and receiving operations in a network and synchronizing clocks as a function of network delay times.

Many methods have been described in the literature to deal with the effect of unreliable transmission on synchronization performance. Most simply, the raw clock offset measurements, denoted as θ_k, can be averaged or filtered by some low-pass. To reduce the effect of delay outliers, the state of the art clips the input to the filter to some empirical 3σ level. For optimum filtering, a known method uses the mathematical clock model to derive a Kalman Filter to estimate the server clock state (phase and frequency) from θ_k. The Kalman filter is optimum for random Gaussian errors on θ_k, but for the case of interest with delay outliers and hence occasional large spikes in θ_k, the Kalman Filter is known to be too sensitive. There is however a vast literature on Robust Kalman Filters which can be applied here: In a known method, if measurements are lost due to network failure, the recursive updating of the state estimate is simply suppressed. In a further known method, the innovation vector in the update step is clipped to some level b, to be chosen in some minimax sense. A known method combines these concepts: if the (norm of the) innovation is larger than a threshold, i.e. an outlier is detected, the updating of the state estimate is suppressed.

The above methods modify recursive filtering to deal with outliers. A known method indicates that non-recursive (batch) procedures may perform better in the presence of outliers. Specifically for synchronization, batch line fitting methods have been described. It is assumed that the clock server periodically broadcasts time messages and considers consecutive differences between client and server time stamps. Delay outliers are then always positive, thus simplifying outlier detection and line fitting. NTP uses the latest eight measurements in an ad-hoc way to detect outliers.

US 2014/068315 A1 relates to methods and devices for clock offset in skew estimation wherein the slave clock employs an independent free-running clock and recursive estimation technique to estimate the clock offset and clock skew between slave and master clocks.

US 2012/0269204 A1 relates to a method for generating a timing correction in timing transfer systems evaluating the weighted average of short-term, medium-term and long-term measurements of local clock off-set.

WO 2015/049480 A1 relates to methods and devices for time and frequency synchronization.

The present invention is an application of batch processing to synchronization, combined with concepts from recursive processing to deal with measurement outliers. A clock server has a high quality clock (master clock) which provides the reference time. A communication link connects the clock server to a client clock (slave clock). The task is to synchronize the client clock to the server clock, i.e. to control the client clock such that the client time x_C (t) closely tracks the server time x_S(t). In the well-known network-based synchronization method (also known as “echo” or “ping-pong” method), client clock offset θ is measured by exchanging timing messages between the server and client clock, and time-stamping the message transmission and reception times T₁ to T₄. This is shown in FIG. 1A. If the frequency error of the client clock is negligible, then the instantaneous clock offset is

$\theta =\frac{T_1-T_2}{2}+\frac{T_4-T_3}{2}+\frac{\Delta d}{2},$

where Δd=d_CS−d_SC is the asymmetry of the message delays from client to server d_CS and server to client d_SC. In practice, these delays are unknown. However it is usually assumed that the bi-directional links are symmetric, i.e. Δd=0, hence

${\theta }_k=\frac{T_1-T_{2}}{2}+\frac{T_4-T_3}{2}$

is the instantaneous measured offset of the client clock (at time t_k=T₁). This formula is well-known and is the basis for the standardized synchronization protocols NTP [RFC 5905] and PTP [IEEE 1588].

The problem to be solved is to mitigate these effects by appropriate processing of the sequence of measurements {θ_k}, and thus to improve performance of network-based synchronization. Hence, it is the object of the present invention to provide an improved synchronization method in a network.

The problem is solved by a method for network-based synchronization of a client clock to a server clock holding features according to claim 1.

Accordingly, a method for synchronization of a client clock to a server clock over a network is provided comprising the steps of determining on the network at least one clock offset, preferably a plurality of clock offsets, by exchanging time synchronization messages between the client and the server clock, wherein the clock offsets are only considered within a provided time horizon. The suggested method furthermore comprises weighting the determined clock offsets by their measurement accuracy, in order to obtain a weighted clock offset that may include an estimated clock phase and clock frequency. The weighted clock offset can be the output of a filter, e.g. using Kalman filter to produce estimated clock phase and clock frequency from a sequence of the determined clock offsets, or preferably the weighted clock offset can be the clock phase and the clock frequency obtained by moving horizon line fitting using equation (3) as described later. After weighting of the clock offsets, the client clock can be adjusted as a function of the weighted clock offset such that the client clock is synchronized to the server clock. As the synchronization is performed over a network, the suggested subject-matter of the independent claims can be accomplished network-based.

The person skilled in the art appreciates that several approaches for determining the at least one clock offset can be applied such as calculation, estimation, measuring or reading-out of such values.

Although a client-server-architecture is the typical application scenario of the present invention, the person skilled in the art appreciates that the involved actors can as well behave in a peer-to-peer fashion. Hence, in the following, the server machine as well as the client machine may both be provided by independent peers, which again can be implemented as hardware resources or software components.

Network-based synchronization comprises the exchanging of timing messages such as timing request messages and/or timing response messages, also known as sync and delay measurement messages. The person skilled in the art appreciates that the timing messages may comprise delay measurements and therefore delay measurement messages can be exchanged. Furthermore, the timing messages to be exchanged can be referred to as synchronization messages or time synchronization messages used in NTP, PTP and other similar protocols. According to IEEE 1588, the PTP message is the general term for synchronization messages, where the most relevant are: Sync, Delay_Req, Follow_Up, Delay_Resp. Regarding the time synchronization according to IEEE 1588, Sync is sent at t1 and Delay_Req at t3.

For each sending operation of such a timing message and for each receiving operation, a time stamp is provided. Hence, it is possible to measure delay times of a network, which means the transmission time between sending and receiving a message. This procedure can be performed repeatedly such that several measurements are taken and therefore the latency time of a network can be evaluated more accurately. It may also be of advantage to send messages according to different properties over the network. For instance, one may measure transmission times of very large messages and compare those evaluation results with transmission times of a variety of very short messages. Hence, the overall network behaviour can be determined.

As only the current network behaviour is of relevance, a time horizon can be determined. Hence, only a specific number of messages or a specific time span or time interval is considered for evaluating the network parameters. As this procedure can be performed several times, at least one clock offset is determined, which is determined as a function of at least one transmission time of at least one message.

Once the at least one clock offset is determined or measured, it is respectively weighted. This can be performed by selecting a provided weighting factor, such as the inverse of the measurement variance.

Once the network behaviour is properly detected, the client clock can be adjusted as a function of the weighted at least one clock offset. This may be performed by the addition of delay times to the client timing.

In the following, further information is provided regarding the technical background of the present invention. The person skilled in the art appreciates that the following mathematical background and aspects of the present invention only provide aspects of the technical teaching and do not limit the scope of the subject-matter. Especially, it may be of advantage to amend the provided technical teaching by additional factors or parameters. While the provided mathematical models are clearly defined by mathematical formulae, the person skilled in the art appreciates that those formulae may as well be enhanced or adapted towards a specific application scenario. Hence, only the basic functionality is provided. Therefore, basic aspects of the present invention are described in the following such as:

• • Aspect A: Clock Model and Kalman Filter for Synchronization
  • Aspect B: Moving Horizon Estimation for Synchronization
  • Aspect C: Simplification of Moving Horizon Estimation to Moving Horizon Line Fitting

Aspect A: Clock Model and Kalman Filter for Synchronization

Clock Model: In the simplest clock model, time deviation x(t) increases linearly in time, i.e.

x(t)=y₀t+x₀  (A.1),

where y₀ is the constant fractional frequency offset. Hence the clock phase x(t_k) at discrete times t_k satisfies the recursion

x(t_k+1)=x(t_k)+y₀(t_k+1−t_k)  (A.2).

More generally, models of imperfect clocks describe clock phase x and frequency deviation y by stochastic processes X and Y, hence

X _k+1 =X _k +Y _k(t_k+1−t_k)  (A.3).

It can be assumed that the clock phase X(t) and frequency deviation Y(t) can be described by independent Brownian motions with intensities σ₁² of (in units of [s]) and σ₂² (in units of [1/s]), respectively, i.e.

dX(t)=(Y(t)+y₀)dt+σ₁dW₁(t)  (A.4),

dY(t)=a dt+σ₂dW₂(t)  (A.5).

If non-constant frequency drift is included, the model includes a third Brownian motion dW₃(t) and becomes three-dimensional.

By integration, a time-discrete linear stochastic clock model is derived: The clock state vector is two-dimensional, X_k=[X_k Y_k]′, and obeys the recursion

$\begin{array}{cc}\left[\begin{array}{c}{X}_{k+1}\\ Y_{k+1}\end{array}\right]=\underset{\underset{F_k}{}}{\left[\begin{array}{cc}1& t_{k+1}-t_k\\ 0& 1\end{array}\right]}\underset{\underset{x_k}{}}{\left[\begin{array}{c}{X}_k\\ Y_k\end{array}\right]}+\underset{\underset{w_k}{}}{\left[\begin{array}{c}{W}_{1,k}\\ W_{2,k}\end{array}\right]}& \left(A\mathrm{.6}\right)\\ x_{k+1}=F_kx_k+w_k,& \left(A\mathrm{.7}\right)\end{array}$

where W_k are i.i.N(0, Q) with

$\begin{array}{cc}Q\stackrel{\mathrm{def}}{=}E\left[w_kw_k^{\prime }\right]=\left[\begin{array}{cc}{\sigma }_1^2\left(t_{k+1}-t_k\right)+{\sigma }_2^2\frac{{\left(t_{k+1}-t_k\right)}^3}{3}& {\sigma }_2^2\frac{{\left(t_{k+1}-t_k\right)}^2}{2}\\ {\sigma }_2^2\frac{{\left(t_{k+1}-t_k\right)}^2}{2}& {\sigma }_2^2\left(t_{k+1}-t_k\right)\end{array}\right].& \left(A\mathrm{.8}\right)\end{array}$

The clock parameters σ₁² and σ₂² are obtained from measurements (data sheets) of the oscillator on which the clock is based, see Note 1 below.

Clock measurement for synchronization: At the client clock side, the server clock phase X_k is measured using the two-way echo measurement involving exchange of timestamps T₁ to T₄. The resulting offset measurement θ_k corresponds directly to the phase X_k. Measurement errors are due to jitter in the time stamping and Packet Delay Variation (PDV) and delay asymmetries Δd of the message transmissions. In the model, these errors are all subsumed in the measurement noise v_k. The measurement equation of the server clock state x_k is thus

$\begin{array}{cc}\underset{\underset{z_k}{}}{\left[{\theta }_k\right]}=\underset{\underset{H^{\prime }}{}}{\left[\begin{array}{cc}1& 0\end{array}\right]}\underset{\underset{x_k}{}}{\left[\begin{array}{c}{X}_k\\ Y_k\end{array}\right]}+\underset{\underset{v_k}{}}{\left[v_k\right]},& \left(A\mathrm{.9}\right)\\ z_k=H^{\prime }x_k+v_k,& \left(A\mathrm{.10}\right)\end{array}$

where the measurement z_k is the scalar θ_k and v_k is the measurement noise with variance

$\begin{array}{cc}{R}_k\stackrel{\mathrm{def}}{=}E\left(v_kv_k^{\prime }\right)={\sigma }_{\theta }^2& \left(A\mathrm{.11}\right)\end{array}$

(in units of [s²]). Occasional large errors in the measurements θ_k, e.g. due to message loss or network congestion, are modelled as large σ_θ²→∞ leading to “additive outliers” in the measurements z_k, i.e. R_k→∞.

Note 1: The values of the clock parameters σ₁² and σ₂² can be obtained from off-line measurements of the Allan variance σ_y²(τ),

$\begin{array}{cc}{\sigma }_y^2\left(\tau \right)\stackrel{\mathrm{def}}{=}\frac{1}{2}〈{\left({\stackrel{\_}{Y}}_{i+1}-{\stackrel{\_}{Y}}_i\right)}^2〉,\mathrm{where}{\stackrel{\_}{Y}}_i=\frac{1}{\tau }{\int }_{i\tau }^{\left(i+1\right)\tau }y\left(s\right)\mathrm{ds}=\frac{x\left(\left(i+1\right)\tau \right)-x\left(i\tau \right)}{\tau },& \left(A\mathrm{.12}\right)\end{array}$

y(t) and x(t) are the measured instantaneous frequency and clock deviation, and τ is the measurement interval. Using the relationship derived in Cristina Zucca and Patrizia Tavella, “The Clock Model and Its Relationship with the Allan and Related Variances,” IEEE Transactions on ultrasonics, ferroelectrics, and frequency control, vol. 52, no. 2, February 2005, pp. 289-296,

$\begin{array}{cc}{\sigma }_y^2\left(\tau \right)=\frac{{\sigma }_1^2}{\tau }+\frac{{\sigma }_2^2\tau }{3},& \left(A\mathrm{.13}\right)\end{array}$

(valid for a=0), σ₁² and σ₂ ² are read off from a plot of the measured Allan variance σ_y²(·): The value of σ_y²(τ) at some small τ gives σ₁², while the value at a large τ gives σ₂².

Note 2: Typical values for PC clocks are σ₁²=10⁻¹² s and σ₂²=10⁻¹⁶ s⁻¹. After a time interval T=1 s, this results in a clock phase variance of σ₁²T=10⁻¹² s², i.e. a jitter of 10⁻⁶ s=1 μs. The fractional frequency variance is σ₂²t=10⁻¹⁶, i.e a standard deviation of 10⁻⁸=0.01 ppm or 0.01 μs/s.

Kalman Filter: Given the clock offset measurements z_k, k=1,2, . . . , the Kalman filter produces minimum mean square error estimates {circumflex over (x)}_k of the state x_k+1, i.e. of both the clock phase (time offset) and frequency offset (see Note 3 below). For the above model, the Kalman filter recursion is

$\begin{array}{cc}{K}_k=F_kP_{k-1}{H\left(H^{\prime }P_{k-1}H+R_k\right)}^{-1}& \left(A\mathrm{.14}\right)\\ {\hat{x}}_k=F_k{\hat{x}}_{k-1}+K_k\left(z_k-H^{\prime }{\hat{x}}_{k-1}\right)& \left(A\mathrm{.15}\right)\\ \begin{array}{c}{P}_k=F_k\left[P_{k-1}-P_{k-1}{H\left(H^{\prime }P_{k-1}H+R_k\right)}^{-1}H^{\prime }P_{k-1}\right]F_k^{\prime }+Q=\\ =\left(F_k-K_kH^{\prime }\right)P_{k-1}\left(F_k-K_kH^{\prime }\right)+K_kR_kK_k^{\prime }+Q,\end{array}& \left(A\mathrm{.16}\right)\end{array}$

with initializations

{circumflex over (x)} ₀ =E[x ₀]  (A.17)

P ₀ =E[(x₀−{circumflex over (x)}₀)(x₀−{circumflex over (x)}₀)′]  (A.18).

The Kalman Filter is mean-square optimal if the measurement noise v_k is uncorrelated with known finite variance R_k. However, any measurement outliers will lead to large errors in the estimates.

To make Kalman filters more robust, typically outliers are detected by comparing the “innovation” signal (Z_k−H′{circumflex over (x)}_k−1 in (A.15) against some threshold. If the threshold is exceeded, an outlier is declared, i.e. the corresponding measurement variance is set to R_k→∞, causing K_k→0, thus the measurement z_k (=θ_k) is ignored and the error covariance P_k increases by Q. The choice of the threshold is critical and may require some ad-hoc tuning.

Client clock control: {circumflex over (x)}_k in (A.15) is the server clock state x_k+1 at time t_k+1, predicted at time t_k based on measurements up to z_k, i.e. {circumflex over (x)}_k={circumflex over (x)}_k+1|k. The estimate of x_k is {circumflex over (x)}_k|k=F_k⁻¹{circumflex over (X)}_k+1|k, i.e. {circumflex over (X)}_k|k={circumflex over (X)}_k+1|k−Ŷ_k|k(t_k+1−t_k) and Ŷ_k|k=Ŷ_k+1|k. For the synchronization of the client clock, these values are to be applied as set points to adjust the client clock at time t_k.

Aspect B: Moving Horizon Estimation (MHE) for Synchronization

Given the measurements z₁, . . . z_N over the horizon k=1, . . . N, the least squares state estimates {circumflex over (x)}₁ . . . {circumflex over (x)}_N minimize

J(x₁, . . . x_N)=Σ_k=1^N(z_k−H′x_k)′R_k⁻¹(z_k−H′x_k)+w′_kQ⁻¹w_k  (B.1),

subject to that the sequence {x_k} is the solution of the model (A.7) when driven by the sequence {w_k}. Define [x₁ . . . x_N]′ and W=[x₁ w₁ . . . w_N−1]′ (vectors of dimension 2N), then J(X) can be written as

J(X)=const+X′X+2f′X+W′W  (B.2)

where

$\begin{array}{cc}ℝ=\left[\begin{array}{ccc}\ddots & 0& 0\\ 0& {\mathrm{HR}}_k^{-1}H^{\prime }& 0\\ 0& 0& \ddots \end{array}\right],f=-\left[\begin{array}{c}⋮\\ {\mathrm{HR}}_k^{-1}z_k\\ ⋮\end{array}\right],\mathbb{Q}=\left[\begin{array}{ccc}0& 0& 0\\ 0& Q^{-1}& \dots \\ 0& ⋮& \ddots \end{array}\right].& \left(B\mathrm{.3}\right)\end{array}$

The least square state estimate is found by solving the Quadratic Programming (QP) problem

$\begin{array}{cc}\underset{X,W}{\min }\left\{X^{\prime }ℝX+2f^{\prime }X+W^{\prime }\mathbb{Q}W\right\}& \left(B\mathrm{.4}\right)\end{array}$

where X, W are subject to the model state equation (A.7).

Note 3: Since H′=[1 0], the 2×2 sub-matrices HR_k⁻¹H′ in , and hence , are singular (positive semidefinite, rather than positive definite). Minimization of X′X+2f′X may thus become an ill-conditioned problem. However, the additional term W′W with non-singular positive definite stabilizes the problem (B.10).

The following steps formulate the QP problem in terms of transformed optimization variables W such that the model state equations are satisfied. Solving the recursion (A.7) gives

x _k=(π_i=1^k−1 F_i)x₁+Σ_i=1^k−1(π_j=i+1^k−1 F_j)w_i  (B.5).

Define

$\begin{array}{cc}{\Phi }_{\mathrm{ki}}\stackrel{\mathrm{def}}{=}\prod _{j=i}^{k-1}F_j=\left[\begin{array}{cc}1& t_k-t_i\\ 0& 1\end{array}\right]& \left(B\mathrm{.6}\right)\end{array}$

where the second equality follows from the definition of F_k. Then

x _k=Φ_k1x₁+Σ_i=1^k−1 φ_k,i+1w_i  (B.7),

or, in explicit form,

$\begin{array}{cc}\left[\begin{array}{c}{X}_k\\ Y_k\end{array}\right]=\left[\begin{array}{c}{X}_1+Y_1\left(t_k-t_1\right)\\ Y_1\end{array}\right]+\sum _{i=1}^{k-1}\left[\begin{array}{c}{W}_{1,i}+W_{2,i}\left(t_k-t_{i+1}\right)\\ W_{2,i}\end{array}\right].& \left(B\mathrm{.8}\right)\end{array}$

In terms of the optimization variable

W=[x ₁ w ₁ . . . w _N−1 ]′=[X ₁ Y ₁ W _1,1 W _2,1 . . . W _1,N−1 W _2,N−1]′,

X thus becomes

$\begin{array}{cc}X=\underset{\underset{\Sigma }{}}{\left[\begin{array}{cccc}{\Phi }_{11}& 0& 0& \dots \\ {\Phi }_{21}& {\Phi }_{22}& 0& \dots \\ {\Phi }_{31}& {\Phi }_{32}& {\Phi }_{33}& \dots \\ ⋮& ⋮& ⋮& \ddots \end{array}\right]}\underset{\underset{W}{}}{\left[\begin{array}{c}{X}_1\\ W_1\\ W_2\\ ⋮\end{array}\right]}=\Sigma W.& \left(B\mathrm{.9}\right)\end{array}$

The Quadratic Programming problem then becomes

$\begin{array}{cc}\underset{W}{\min }\left\{W^{\prime }\left({\Sigma }^{\prime }\mathrm{ℝ\Sigma }+\mathbb{Q}\right)W+2\left(f^{\prime }\Sigma \right)W\right\}& \left(B\mathrm{.10}\right)\end{array}$

with bound constraints (see below)

|W|≦q  (B.11).

The solution W is obtained by some standard numerical QP algorithm. The estimated state trajectory {circumflex over (x)}₁ . . . {circumflex over (x)}_N is recovered using X=ΣW. In practice, only the most recent estimate {circumflex over (x)}_N is of interest. It is applied to the control of the client clock at time k=N. The estimation horizon k=1, . . . N is then moved one or T steps forward.

Constrained estimation due to measurement outliers: Consider the clock phase change X_k+1−X_k. Its component which is not due to the frequency deviation Y_k is W_1,k. As W_1,k˜N(0, q₁₁), then with high probability,

$\begin{array}{cc}W_{1,k}\le m\sqrt{q_{11}}=m\sqrt{{\sigma }_1^2\left(t_{k+1}-t_k\right)+{\sigma }_2^2\frac{{\left(t_{k+1}-t_k\right)}^3}{3}}& \left(B\mathrm{.12}\right)\end{array}$

(using (A.8)), where m is the ratio of the maximum over the standard deviation, e.g. m=2.—The frequency deviation Y_k is modelled as a random walk with i.i.d. increments W_2,k˜N(0, q₂₂). With high probability

|Y_k+1−Y_k|=|W_2,k|≦m√{square root over (q₂₂)}=m√{square root over (σ₂²(t_k+1−t_k))}  (B.13)

(see (A.8), i.e. the components of W are bounded. Estimates corresponding to large W_1,k, W_2,k, i.e. with large jumps, are likely artifacts caused by large errors and outliers in the measurements, and should thus be prevented. This is achieved by forcing the estimates to be constrained in |W|≦q.

Outlier detection: In Kalman Filters, outliers are typically detected by observing the “innovation” signal. Using MHE or MHLF described in Aspect C below, no innovation signal is available, and some alternative outlier detection method is required. Here, a simple method is proposed based on observing successive measurements θ_k: From (A.3),

θ_k−θ_k−1=Y_k−1(t_k−t_k−1)+v_k−v_k−1  (B.14).

Unusually large measurement errors v_k are considered as outliers, hence θ_k is classified as an outlier if

|θ_k−θ_k−1−Ŷ_k−1(t_k−t_k−1)|>mσ_θ  (B.15),

where m is the ratio of the threshold to the steady state measurement deviation σ₀. When outliers are declared, then the corresponding measurement value z_k (=θ_k) is suppressed by setting its inverse weight R_k to a large value, e.g. R_k=1000σ₀², in (A.14), (B.1), or (C.3).

Aspect C: Simplification of MHE to Moving Horizon Line Fitting (MHLF)

Assuming W≡0, it follows from (B.8) that the past states x_k, at times t_k, are related to the current state x_k=[X_k Y_k]′ at time t_k by

$\begin{array}{cc}{x}_{k^{\prime }}=\left[\begin{array}{c}{X}_{k^{\prime }}\\ Y_{k^{\prime }}\end{array}\right]=\left[\begin{array}{c}{X}_k-Y_k\left(t_k-t_{k^{\prime }}\right)\\ Y_k\end{array}\right].& \left(C\mathrm{.1}\right)\end{array}$

The squared error (B.1) over the moving horizon of the most recent N measurements at t_k−N+1, . . . t_k becomes

$\begin{array}{cc}J\left(x_{k-N+1},\dots x_k\right)=\sum _{k^{\prime }=k-N+1}^k{\left(z_{k^{\prime }}-H^{\prime }x_{k^{\prime }}\right)}^{\prime }R_{k^{\prime }}^{-1}\left(z_{k^{\prime }}-H^{\prime }x_{k^{\prime }}\right)=\sum _{k^{\prime }=k-N+1}\frac{1}{R_{k^{\prime }}}{\left(z_{k^{\prime }}-X_k+Y_k\left(t_k-t_{k^{\prime }}\right)\right)}^2,& \left(C\mathrm{.2}\right)\end{array}$

using H′=[1 0]. Under the assumption of independent Gaussian measurement errors v_k with known variances R_k, minimization of (C.2) is identical to Maximum Likelihood estimation of X_k and Y_k. This interpretation motivates the choice of the objective function (C.2). Thus minimization of J(.) reduces to simple line fitting, i.e. minimization with respect to the clock offset X_k and frequency offset (slope) Y_k. These are found by solving the normal equation

$\begin{array}{cc}\left[\begin{array}{cc}\sum _{k^{\prime }=k-N+1}^k\frac{1}{R_{k^{\prime }}}& -\sum _{k^{\prime }=k-N+1}^k\frac{\left(t_k-t_{k^{\prime }}\right)}{R_{k^{\prime }}}\\ -\sum _{k^{\prime }=k-N+1}^k\frac{\left(t_k-t_{k^{\prime }}\right)}{R_{k^{\prime }}}& \sum _{k^{\prime }=k-N+1}^k\frac{{\left(t_k-t_{k^{\prime }}\right)}^2}{R_{k^{\prime }}}\end{array}\right]\left[\begin{array}{c}{X}_k\\ Y_k\end{array}\right]=\hspace{1em}\left[\begin{array}{c}\sum _{k^{\prime }=k-N+1}^k\frac{z_{k^{\prime }}}{R_{k^{\prime }}}\\ -\sum _{k^{\prime }=k-N+1}^k\frac{z_{k^{\prime }}\left(t_k-t_{k^{\prime }}\right)}{R_{k^{\prime }}}\end{array}\right].& \left(C\mathrm{.3}\right)\end{array}$

For increasing horizons, i.e. N=k, the well-known Recursive Least Squares (RLS) algorithm would allow to recursively compute X_k, Y_k, but in terms of ever increasing factors (t_k−t₁) which is however impractical.

For a moving horizon of constant size N, there are no recursive algorithms. The equation (C.3) must be set up and solved at each step. For constant update intervals T, i.e. t_k=kT, the elements in (C.3) can be computed by N-tap FIR filters: for example, the second element of the RHS of (C.3) is the output of an FIR filter with coefficients f_i=iT, i=0 . . . N−1, and input −z_k/R_k, and similarly for the other elements of (C.3).

According to a further aspect of the present invention, the clock offset z_k=θ_k is measured as a function of a transmission time measured between sending and receiving the timing measurement messages. This provides the advantage that a known method can be reused in the present application scenario and can be furthermore adapted towards the technical teaching of the present invention. Hence, legacy systems can easily be adapted.

According to a further aspect of the present invention, the client determines the at least one clock offset as a function of time stamps, each time stamp indicating a send or receive operation of the time delay measurement messages. This provides the advantage that relevant method steps are performed on the client side, which allows the client to actively perform the suggested method while the server only needs to respond to specific messages.

According to a further aspect of the present invention, the time stamps are based on a local free-running oscillator. This provides the advantage that the client time stamps, for instance time stamps T₁ and T₄ as described with reference to FIG. 2 are based on a local free-running oscillator driving a counter, such as a HW counter. Instead of applying a known closed-loop synchronization, the time stamps are based on a free-running time base, running in open loop. This will also be described in the specific embodiments.

According to a further aspect of the present invention, the step of weighting by measurement accuracy is performed by selecting the inverse measurement variance parameter as a weighting parameter. Furthermore the person skilled in the art recognizes that the step of weighting by the measurement accuracy may also be performed by selecting a measurement variance parameter as an inverse weighting parameter. This provides the advantage that according to the present invention, the specific network behaviour and parameters can be considered for adjusting the client clock.

According to a further aspect of the present invention, in case all sending and receiving operations are performed properly, a default value is provided for the variance parameter. This provides the advantage that the at least one offset can be weighted according to a determined parameter which can be specifically chosen for a specific network scenario.

According to a further aspect of the present invention, in case the sending or the receiving operation is failed, an infinite value is provided for the variance parameter. The infinite value refers to a value which is much larger than the default value. In the specific embodiment, much larger means that a specific factor is provided which can be multiplied with the default value for determining such a large value. For instance, instead of choosing an infinite value, a factor of 1000 can be chosen which is multiplied on the default value for the variance parameter.

According to a further aspect of the present invention, in case one determined clock offset differs from previously measured clock offsets by a specified extent, an infinite value is provided for the variance parameter. The person skilled in the art appreciates that the infinite value which is provided for the variance parameter can again be a value which is much larger than the default value, for instance by means of the definition of infinite as provided above.

According to a further aspect of the present invention, the time horizon is determined by at least one of a group of techniques, the group comprising a number of previous message exchanges and a determined time span. This provides the advantage that a specific time interval, for instance being specified by a number of seconds, can be considered for adjusting the client clock. Therefore, the time horizon can be specified by the last N timing message exchanges. Hence, it is possible to specify that a number of previous message exchanges or a specific time span is considered. This can also be referred to as an expiry time, as older measurements are not considered. One example for such a number N of the last timing message exchanges is a range between 10 and 1.000, namely 10<N<1000. “Moving Horizon” and “sliding window” refer to the “time horizon”. Hence, this terminology can be used interchangeably.

According to a further aspect of the present invention, determining the timing offset considers asymmetric transmission times between the client and the server. This provides the advantage that the transmission time from the client to the server may differ from the transmission time from the server to the client. Again, it is noted that the client and server architecture can be enhanced towards a peer-to-peer architecture, which, again, may provide asymmetric transmission times. Hence, the network behaviour can be measured more precisely.

According to a further aspect of the present invention, one of a group of estimation techniques is applied, the group comprising a Moving Horizon Estimation and a line fitting over a sliding window technique. This provides the advantage that known expiry time techniques can be applied and enhanced towards the technical teaching of the present invention.

According to a further aspect of the present invention, the measurement accuracy is perceived as a function of at least one of a time stamping inaccuracy due to hardware and software delays in the clocks, a message loss in the communication network, and time-varying delays and delay asymmetries due to network load and congestion. This provides the advantage that a variety of error sources can be addressed.

The object is also solved by a system for synchronization of a client clock to a server clock over a network, comprising a determining unit being arranged to determine on the network at least one clock offset by exchanging time synchronization messages between the client and the server, wherein clock offsets are only considered within a provided time horizon. The system also comprises a weighting unit being arranged to weight the determined at least one clock offset by their measurement accuracy and furthermore comprises an adjustment unit being arranged to adjust the client clock as a function of the weighted clock offsets such that the client clock is synchronized to the server clock.

The problem is also solved by a client machine being arranged to perform a synchronization of a client clock to a server clock over a network, comprising a determining unit, a weighting unit, and an adjustment unit which are arranged according to those units as provided above.

The person skilled in the art appreciates that the system comprises a client and a server or respective peers and the suggested client can be operated independently of the server, which means that the client can perform the measurements, the weighting and the adjustment itself. Hence, the server machine only needs to send the timing messages back to the client machine.

Furthermore, a computer-readable medium containing program instructions for causing a computer to perform the provided method is suggested.

This invention gives a simple and robust method for network-based synchronization, with improved performance in the presence of measurement outliers as due to time-varying and lossy network connections. The method is derived in systematic way and is optimum in a well defined sense (as given by the objective functions (B.1) or (C.2)).

The only parameters to be selected for the algorithm are horizon length NT and outlier detection threshold mσ_θ. They have intuitive interpretations and are thus simple to tune.

The invention improves synchronization performance even if the underlying time stamping method has low accuracy and the network connection is time-varying and lossy. This lowers the implementation cost of network-based synchronization, as it may allow software- (rather than hardware-) based time stamping and use of low cost lossy networks (rather than broadband deterministic) networks.

The moving horizon method used in the invention may require that a batch of N measurement points is stored in local memory before processing can start, in contrast to recursive filtering approaches where only the newest clock state needs to be stored and updated.

The time constant of the synchronization may be given by the horizon length, i.e. NT seconds. This affects in particular the synchronization response time when the client oscillator experiences fast changes in clock phase or frequency, e.g. due to vibration or temperature change. NT must be chosen as a compromise between high measurement accuracy and outlier detection reliability, requiring high NT, and low time lag of the control response, requiring low NT.

Note that synchronization accuracy and response time may depend also on the oscillator stability and the implementation of the “gearing” control between the free-running oscillator and synchronized clock at the client.

In the following, the invention will be described merely by way of illustration with reference to the accompanying Figures.

FIG. 1A shows a known synchronization method;

FIG. 1B shows a method for network-based synchronization of a client clock to a server clock according to an aspect of the present invention;

FIG. 2 shows a device being arranged to perform a network-based synchronization according to an aspect of the present invention;

FIG. 3 shows simulation results of the method for network-based synchronization illustrating the improvements being achieved by applying the suggested invention; and

FIG. 4 shows an aspect of a method for network-based synchronization of a client clock to a server clock according to the present invention.

The invention improves the well-known network-based synchronization method shown in FIG. 1. One of the main novelties is to apply Moving Horizon Estimation (MHE) to perform optimum batch estimation of the clock phase and frequency offset. MHE allows to introduce explicit constraints in the estimation, in order to suppress the effect of measurement outliers on the estimates. It also allows to weight measurements, based on their perceived accuracy. Thus estimation of phase and frequency is improved in the presence of outliers. The estimates are used to control the client clock, which results in an improved synchronization performance.

The use of MHE or MHLF requires that the client time stamps (T₁ and T₄) are based on a free-running time base (oscillator), i.e. which runs in open loop. (This is in contrast to traditional closed-loop synchronization using PLLs and coupled oscillators [NTP], where time stamps are based on the synchronized clock which is controlled in closed loop fashion. This would affect subsequent T₁ and T₄ measurements within the estimation horizon, which is undesired). Thus the open loop architecture, shown in FIG. 2, is prerequisite and part of the invention.

The preferred embodiment uses a simplified form of MHE, here called Moving Horizon Line Fitting (MHLF). The client clock may be synchronized to the server clock using the following steps:

• • 1. Raw measurements: At time t_k, the client clock initiates a timing message exchange with the server clock, using the well-known protocol with timing request and timing response. From the resulting time stamps stamps T₁ to T₄, the client clock calculates the instantaneous clock offset θ_k,

$\begin{array}{cc}{\theta }_k=\frac{T_1-T_2}{2}+\frac{T_4-T_3}{2}.& \left(1\right)\end{array}$

• •  According to the invention, the client time stamps (T₁ and T₄) are based on a local free-running oscillator driving a counter (“HW counter” in FIG. 2).
  • 2. Measurement weighting and outlier detection: Measurements can be weighted according to their perceived accuracy, by selecting the “measurement variance” parameter R_k, to be used in (3) below. An example method is as follows: (i) If the message exchange was normal and successful, R_k is set to some small default value R_k=R. (ii) If the message exchange failed, i.e. the transmission of the timing request or response messages was unsuccessful, R_k is set to infinite, or to some large value such as R_k=1000R. (iii) Also, if the measurement is as an outlier, set R_k=1000R. A simple and robust method to detect outliers considers successive raw measurements:
  •  If they differ by much more than expected, given the frequency offset, i.e. if

|θ_k−θ_k−1−Ŷ_k−1(t_k−t_k−1)|>mσ_θ  (2),

• •  then θ_k is classified as an outlier and its corresponding R_k=1000R. In (2), Ŷ_k−1 is the current estimate of the frequency offset, and mσ_θ is a suitably chosen threshold, see (B.15) in Aspect B.
  • 3. Moving Horizon Line Fitting: The client collects the measurements z_k=θ_k over a moving horizon from t_k−N+1 to t_k, as obtained by the last N timing message exchanges. It sets up and solves the linear system of equations (3) to obtain the new estimates of clock phase {circumflex over (X)}_k and frequency Ŷ_k,

$\begin{array}{cc}\left[\begin{array}{cc}\sum _{k^{\prime }=k-N+1}^k\frac{1}{R_{k^{\prime }}}& -\sum _{k^{\prime }=k-N+1}^k\frac{\left(t_k-t_{k^{\prime }}\right)}{R_{k^{\prime }}}\\ -\sum _{k^{\prime }=k-N+1}^k\frac{\left(t_k-t_{k^{\prime }}\right)}{R_{k^{\prime }}}& \sum _{k^{\prime }=k-N+1}^k\frac{{\left(t_k-t_{k^{\prime }}\right)}^2}{R_{k^{\prime }}}\end{array}\right]\left[\begin{array}{c}{\hat{X}}_k\\ {\hat{Y}}_k\end{array}\right]=\hspace{1em}\left[\begin{array}{c}\sum _{k^{\prime }=k-N+1}^k\frac{z_{k^{\prime }}}{R_{k^{\prime }}}\\ -\sum _{k^{\prime }=k-N+1}^k\frac{z_{k^{\prime }}\left(t_k-t_{k^{\prime }}\right)}{R_{k^{\prime }}}\end{array}\right].& \left(3\right)\end{array}$

• •  The derivation of (3) is given in Aspect C. As seen, a high R_k reduces the weight of the corresponding measurement z_k=θ_k.
  • 4. Clock control: {circumflex over (X)}_k and Ŷ_k are the estimate of the server clock phase and frequency, as computed by the client. In order to track the server clock, for client clock control takes these values as setpoints u_x={circumflex over (X)}_k and u_Y=Ŷ_k. These setpoints are used to adjust the client clock, by adjusting the “gearing” between its free-running oscillator (“HW counter” in FIG. 2) and the synchronized clock (“SW clock” in FIG. 2). Methods are known and can be applied to ensure that the resulting time trajectory increases monotonously, even in the presence of possibly negative jumps in the clock phase setpoint u_x={circumflex over (X)}_k.

As described above, MHLF forces the clock phase estimate to be an exact straight line within the measurement horizon, by assuming that the clock state differences w_k≡0. With long horizons required in high noise environments, this may be too restrictive. Non-zero state differences should thus be allowed, but still constrained to some limits |w_k|≦q, in order to suppress or mitigate the effect of outliers on the state estimates. This leads to a constrained estimation problem, implemented by Moving Horizon Estimation (MHE). Application of MHE to synchronization constitutes another embodiment of this invention; it is derived in Aspect B. At each step, a full Quadratic Programming (QP) problem must be set up and solved numerically, which requires much higher computations than the simple MHLF case of (3). However, QP convergence can be speeded up by orders of magnitude by taking the computed solution from the previous step as starting point for the QP iteration in the next step.

Over very long horizons, the mathematical clock model in (A.7) should not neglect the frequency drift and is thus three-dimensional. Conceptionally, it is straightforward to extend Kalman Filtering and MHE to this case. The frequency drift leads to a quadratic deterministic term in the clock offset, and MHLF becomes a Moving Horizon quadratic curve fitting. However, in practice these extensions and possible embodiments may not be relevant for most applications.

In FIG. 3, estimation results of the suggested method according to the present invention are provided. Measurement results of a Steady-State Kalman Filter, a Time-Varying Kalman Filter as well as a Moving Horizon Line Fitting is provided. The Time-Varying Kalman Filter is denoted by reference sign TV-KF and the Moving Horizon Line Fitting is denoted by reference sign MH-LF.

FIG. 3 shows simulation results of clock estimation methods applied to measured data {θ_k}, showing improvement by the Moving Horizon Line Fitting (MHLF) method of the invention over known Kalman Filter (KF) methods.

Measurement data {θ_k}: measurement duration 4500 seconds, cycle time T=1 s. Server and client clocks were connected over Ethernet LAN (measured standard deviation σ_θ=24.8 μs), or wireless LAN (in intervals 700-800 s and 2400-3200 s, with much higher measured σ_θ). FIG. 3 shows the following:

• • dashed curve: “steady-state KF” with constant Kalman gain K_k, no outlier detection
  • TV-KF curve: “time-varying KF” with outlier detection, time-varying K_k, outlier detection threshold on innovation signal=5·10⁻³ s
  • MH-LF curve: MHLF, horizon NT=300·1 s, outlier detection threshold mσ_θ=2·10⁻⁴ s. One example for a number N of the last timing message exchanges is a range between 10 and 1.000, namely 10<N<1000, preferably N=300.
  • top: clock phase {θ_k}, and its estimates as KF output, and MHLF output
  • center: clock phase estimates with overall slope removed, to show differences in clock estimation errors
  • bottom: clock frequency estimates.

Note that the maximum estimation error in the presence of large outliers is smallest for MHLF.

FIG. 4 shows an aspect of a method for network-based synchronization of a client clock to a server clock according to the present invention. In a first method step, a timing measurement message, also referred to as time synchronization message or synchronization message throughout the description, is sent from the client to the server 100. The sending operation as well as further receiving operations have one time stamp each which is assigned to the respective sending or receiving time. The sent message is received by the server 101. These method steps 100 and 101 may be performed repeatedly, for instance, several messages may be sent from the client to the server and several messages may be again sent from the server to the client, for instance according to a sending profile. Hence, it is possible to send messages according to a predefined time schedule. Each of the transmission times from client to server and server to client is determined and/or stored in a method step 102. In a further method step 103, it is evaluated whether the performed measurements are within the provided time horizon. Hence, message step 103 checks the respective time stamps and compares them with the provided time horizon. Only in case the measurements are within the time limit as being defined by the time horizon, the further method step 104 of weighting the measured clock offsets of step 102 is performed. After a weighting of the measured clock offsets is performed, the adjustment setpoints are calculated according to the described methods and the client clock can be adjusted and/or corrected accordingly in a further method step 105. The person skilled in the art appreciates that all of the method steps may be performed iteratively and/or in a different order wherever applicable.

Along with the suggested devices, systems, and modules, respective methods for their operation are provided as well as a computer-readable medium having stored thereon instructions executable by a computer processor, the instructions, which, when executed by the processor, perform the method of the aspects as set forth above.

While the invention has been described in detail in the drawings and foregoing description, such description is to be considered illustrative or exemplary and not restrictive. Variations to the disclosed embodiments can be understood and effected by those skilled in the art and practising the claimed invention, from a study of the drawings, the disclosure, and the appended claims. In the claims, the word “comprising” does not exclude other elements or steps, and the indefinite article “a” or “an” does not exclude a plurality. The mere fact that certain elements or steps are recited in distinct claims does not indicate that a combination of these elements or steps cannot be used to advantage, specifically, in addition to the actual claim dependency, any further meaningful claim combination shall be considered disclosed.

Claims

1. A method for synchronization of a client clock to a server clock over a network, comprising: determining, within a provided time horizon, clock offsets by exchanging synchronization messages between the client and server clock;weighting the clock offsets by their measurement accuracy to obtain a weighted clock offset; andadjusting the client clock as a function of the weighted clock offset such that the client clock is synchronized to the server clock,wherein the weighted clock offset includes an estimated clock phase and clock frequency, andwherein the time horizon applies to one of a group of estimation techniques, the group comprising a Moving Horizon Estimation or line fitting over a sliding window technique.

2. The method according to claim 1, wherein the clock offsets are determined as a function of a transmission time measured between sending and receiving the synchronization messages.

3. The method according to claim 1, wherein the client clock determines the clock offset as a function of time stamps, each time stamp indicating a send or receive operation of the synchronization messages.

4. The method according to claim 3, wherein the time stamps are based on a local free-running oscillator.

5. The method according to claim 1, the step of weighting by the measurement accuracy is performed by selecting a measurement variance parameter as an inverse weighting parameter.

6. The method according to claim 5, wherein in case all sending and receiving operations are performed properly a default value is provided for the variance parameter.

7. The method according to claim 1, wherein in case the sending or the receiving operation is failed an infinite value is provided for the variance parameter.

8. The method according to claim 1, wherein in case one determined clock offset differs from previously determined clock offsets to a specified extend an infinite value is provided for the variance parameter.

9. The method according to claim 1, wherein the time horizon is determined by at least one of a group of techniques, the group comprising a number of previous message exchanges, preferably the last N messages, and a determined time span.

10. The method according to claim 1, wherein determining the timing offset considers asymmetric transmission times between the client and the server clock.

11. The method according to claim 1, wherein the measurement accuracy is perceived as a function of at least one of a time stamping inaccuracy due to hardware and software delays in the clocks, a message loss in the communication network, and time-varying delays and delay asymmetries due to network load and congestion.

12. A system for synchronization of a client clock to a server clock over a network, comprising: a determining unit being arranged to determine, within a provided time horizon, at least one clock offset by exchanging synchronization messages between the client and the server lock;a weighting unit being arranged to weight the determined at least one clock offset by their measurement accuracy, in order to obtain a weighted clock offset including an estimated clock phase and clock frequency; andan adjustment unit being arranged to adjust the client clock as a function of the weighted clock offset such that the client clock is synchronized to the server clock,wherein the time horizon applies to one of a group of estimation techniques, the group comprising a Moving Horizon Estimation or line fitting over a sliding window technique.

13. A client being arranged to perform a synchronization of a client clock to a server clock over a network, comprising: a determining unit being arranged to determine, within a provided time horizon, at least one clock offset by exchanging synchronization messages between the client and the server clock;a weighting unit being arranged to weight the determined clock offsets by their measurement accuracy, to obtain a weighted clock offset including an estimated clock phase and clock frequency; andan adjustment unit being arranged to adjust the client clock as a function of the weighted clock offset such that the client clock is synchronized to the server clock, wherein the time horizon applies to one of a group of estimation techniques, the group comprising a Moving Horizon Estimation or line fitting over a sliding window technique.

14. A computer-readable medium containing program instructions for causing a computer to perform the method of claim 1.

15. The method according to claim 2, wherein the client clock determines the clock offset as a function of time stamps, each time stamp indicating a send or receive operation of the synchronization messages.

16. The method according to claim 15, the step of weighting by the measurement accuracy is performed by selecting a measurement variance parameter as an inverse weighting parameter.

17. The method according to claim 16, wherein in case the sending or the receiving operation is failed an infinite value is provided for the variance parameter.

18. The method according to claim 17, wherein in case one determined clock offset differs from previously determined clock offsets to a specified extend an infinite value is provided for the variance parameter.

19. The method according to claim 18, wherein the time horizon is determined by at least one of a group of techniques, the group comprising a number of previous message exchanges, preferably the last N messages, and a determined time span.

20. The method according to claim 19, wherein determining the timing offset considers asymmetric transmission times between the client and the server clock.