Packet TDEV, MTIE, and MATIE

- for Estimating the Frequency and Phase Stability of a

Packet Slave Clock

Antti Pietiläinen

Soc Classification level

1 © Nokia Siemens Networks Antti Pietiläinen / 5.11.2008
Expressing performance of clocks

TIE = Time interval error

MTIE = Maximum time interval error
MTIE/ = Maximum frequency error estimate from MTIE
TDEV = Time deviation
MDEV = Modified Allan deviation
MATIE = Maximum average time interval error (proposal)
MAFE = Maximum average frequency error (proposal)

Antti Pietiläinen / 5.11.2008

Time interval error (TIE)

• Indicates the phase of a clock compared with the phase of a reference

clock. For example if the frequency of a clock is 10 ppb too large, TIE will
increase by 10 ns every second.
• The quantities considered of interest in standardization bodies for
characterization of time and frequency stability are calculated from TIE,
e.g. MTIE, TDEV, and MDEV.

0.0002
TIE [s]

0.0000 TIE of a packet clock

synchronized over a
DSL connection.

-0.0002
50000 100000 150000 200000 250000 300000 350000
Time [s]

Antti Pietiläinen / 5.11.2008

MTIE - maximum time interval error
• MTIE defines maximum wander within an observation window. The
observation window is slid over the TIE data.
• The sizes of the observation windows are indicated by the x-axis values of the
MTIE curve.
• MTIE is specified in G.810. MTIE masks have been specified, for example, in
G.812, G.813, G.823, G.824, and G.8261
T = (N - 1) 0
x(t)

 = n0
Time
error

xppk

(from G.810)

0 1 2 3 k k+n N i
Antti Pietiläinen / 5.11.2008 T1308800-96
MTIE of a practical packet clock
• Each observation window scanned over the TIE curve (see previous slide) will
produce one point on the MTIE or MRTIE curve.
• MTIE of a TCXO based packet clock synchronized over a production network and
DSL connection shoots about one order of magnitude over the PDH specifications.
• Does one need to average this clock for 10000 seconds to go below 16 ppb?
1E-3

MTIE of a packet clock

1E-4
MTIE of clock synchronized over a
DSL connection.
MTIE [s]

G.823 2 Mbit/s
traffic interf. G.824 1.5 Mbit/s
1E-5
network interf.

1E-6 15 ppb
16 ppb OBSAI RP1

1E-7
0.1 1 10 100 1000 10000 100000
tau [s]
Antti Pietiläinen / 5.11.2008
MTIE is a pessimistic estimator when used to
indicate achievable frequency stability

• MTIE predicts a 1000 s

stability of 127 ppb.
60.0µs
• However, by passing
the signal through a 40.0µs MTIE(1000 s)
1st order low-pass =127 µs.
filter with time Original TIE
20.0µs MTIE/ = 127 ppb
constant of 1000 s,
fmax= 500 ppb
TIE [s]

the maximum 0.0s

frequency error
MTIE(1000 s)
drops to 67 ppb. -20.0µs = 12.8 µs.
• With a better filter, MTIE/  = 12.8 ppb
the frequency error -40.0µs fmax= 67 ppb
could be further
After passing low-pass filter
reduced without -60.0µs
of time constant = 1000 s
widening the window.
0 2000 4000 6000 8000 10000 12000
Time [s]

Antti Pietiläinen / 5.11.2008

 Time deviation (TDEV)

• TDEV indicates phase

variation of a clock as a 1E-4

function of averaging
time. 1E-5
• TDEV masks have been
specified in various ITU-

TDEV [s]
T recommendations. 1E-6

• TDEV averages out

extremes. 1E-7

• TDEV and MDEV (on TDEV of a packet clock

following slide) are 1E-8
synchronized over a
closely related, see DSL connection.
formula.
1E-9
1 10 100 1000 10000 100000
3 TDEV
MDEV  tau [s]


Antti Pietiläinen / 5.11.2008

Modified Allan deviation (MDEV)
(root of modified Allan variance)
• Is used to indicate frequency variation and frequency uncertainty of clocks.
• Because frequency variation is built up from various noise phenomena, the
fundamental accuracy limits of a clock can be determined by averaging the clock
frequency over a period of time. Consequently, Allan variance is indicated in graphs as
a function of averaging window size.
• Allan variance is an averaging function and hides occasional bad performance.
Modified Allan deviation

1E-8
16 ppb

1E-9 MDEV of a packet clock

synchronized over a
DSL connection.

1E-10
1 10 100 1000 10000 100000
Averaging window, tau [s]

Antti Pietiläinen / 5.11.2008

MATIE (maximum average time interval error),
MAFE (maximum average frequency error)
• A new metric has to be defined.
• The target: a metric that
– Describes the upper bound of the phase or frequency error of a clock.
– Is presented as a function of averaging window width, similar to TDEV
or MDEV. The averaging window size at which the metric drops below
required level corresponds to the required local oscillator stability.
– Can be calculated from packet delay of fastest packets and from the
phase (TIE) of a clock.
• Solution :
– Calculate average phase (or packet delay) difference between two
windows next to each other.
– Find the maximum value of the difference over the whole data.

Antti Pietiläinen / 5.11.2008

Calculating Maximum average time interval error
• Calculate xi = abs[average(blue)-average(purple)]
• Slide the windows over the data and find max(xi)
• Change the window size and start over until all window sizes of interest have
been covered.

MATIE can be
calculated from time
interval error (TIE) of a
clock, as well as from
packet delays. In case
of packet delays, a
percentile selection is
done first.
In the figure, 10
fastest of each 1000
consecutive samples
have been pre-
selected and
averaged.

Antti Pietiläinen / 5.11.2008

Determining the maximum average time interval error MATIE
and max. average frequency error MAFE at  = 3000s.
TIE (time interval error) is averaged over observation window and maximum
change between two consecutive windows is determined.

0.02292 Note, this is a simplification. In reality the adjacent averaging windows slide over the data.

0.02290 The example shows (Max average freq.

err. MAFE =
how MATIE
MATIE  = MAFE
atand 3000 s MATIE/ =18 ppb)
The function
0.02288 are calculated at estimates
observation window of achievable
0.02286 3000 s.  = 3000 s worst-case
Average TIE [s]

The metric is calculated stability over

TIE [s]

0.02284
at different window the whole TIE
0.02282 sizes. MATIE and MAFE measurement
are plotted as a function MATIE as a function
0.02280 of observation window =53 µs of the
MATIE = 53 µs
size. averaging
0.02278 time of the
clock
0.02276 algorithm.

0.02274
0 10000 20000 30000 40000
Time [s]
Antti Pietiläinen / 5.11.2008
Comparison between frequency stability
estimations of MDEV, MTIE/, and MAFE
(corresponds to TDEV, MTIE, and MATIE in phase domain)

MDEV, MAFE, max freq error, MTIE/ [relative]

• MDEV varies
depending on fill 1E-6
1st order low-pass filtered maximum
ratio and is frequency error. Time constant = tau
optimistic. 80µs

• MTIE/ is too 1E-7

40µs

pessimistic. TIE
0s

MTIE/
• The different metrics
-40µs

-80µs

estimate very 1E-8 0 2000 4000 6000

Time [s]
8000 10000 12000 16 ppb
different frequency 100.0µs

stability. 80.0µs

60.0µs

MDEV
• It is expected that
40.0µs

20.0µs
TIE [s]

1E-9 0.0s

low-pass filtered -20.0µs

-40.0µs

max. frequency -60.0µs MAFE

-80.0µs

error will approach

-100.0µs
0 2000 4000 6000 8000 10000 12000

1 Time [s] 10 100 1000 10000

MAFE when the tau [s]
filter is more
optimal.
Antti Pietiläinen / 5.11.2008
MATIE follows quite closely the framework set by
TDEV and MTIE
n k 1
1
MATIE n 0   max
1k  N 2 n1 n
 x
i k
in  xi  , n = 1, 2, ..., integer part (N/2)

2
N 3n 1 n  j 1 
TDEVn 0     xi  2 n  2 xi  n  xi  , n = 1, 2, ..., integer part  N 
1
6n 2  N  3n  1
 
   3
j 1 i j

 
MTIE(n 0 )  max  max xi  min xi  , n  1,2,... , N  1
1 k  N n k i  k n k i  k n 

n0 is the observation window length, n is the number of samples in the window, 0
is the sample interval, N is the number of samples in the data set. Index variable i is
incremented to scan across the window and k or j is incremented for sliding the
window.

Antti Pietiläinen / 5.11.2008

Estimating packet clock performance from packet
delay variation
Packet delay variation is not a good indicator of expected packet clock properties
because the time interval error of a packet clock is a small and unknown fraction of
the packet delay variation.

Black: Timing packet

delay
Yellow: TIE of the
packet clock

Antti Pietiläinen / 5.11.2008

Describing packet networks’ capability to support
packet timing: Proposals
Measurement Author ‘Published’ Time

Minimum TDEV Symmetricom ITU contrib. June 2007

accepted
Minimum picking Semtech ITU contrib. June 2007
TDEV
Percentile TDEV Symmetricom No public or Taken into
standards discussion in
documents ~December 2007
Fixed selection Nokia Siemens Not published
window percentile Networks
TDEV & MDEV
Fixed selection Nokia Siemens ITU contrib. May 2008
window percentile Networks
MTIE
MATIE, MAFE Nokia Siemens ITU contrib. Sep 2008
Networks

Antti Pietiläinen / 5.11.2008

Comparing modified Allan deviation of packet
clock and packet MDEV of delay values
• At small averaging windows the short-term stability of the local oscillator
dominates the packet clock performance.
• At certain point the curves almost combine.
• TDEV (and MDEV) first showed the link between packet delay variation and
packet clock performance!
1E-4

1E-5 Packet delay MDEVs

Modified Allan deviation

1E-6 0.1 % … 1 % fastest

packets selected from
1E-7 various window sizes

1E-8 16 ppb

1E-9
Packet clock MDEV
1E-10

1E-11
1 10 100 1000 10000 100000
Averaging window, tau [s]
Antti Pietiläinen / 5.11.2008
Packet MTIE vs. packet clock MTIE
• Packet MTIE would allow setting a performance limit mask.
• However, in addition to being pessimistic considering frequency stability, it is
difficult to find parameters for packet MTIE that would accurately describe the
performance of a packet clock.

1E-3 1/100 Minimum picking (6-s window)

1/1000 Minimum picking (60-s window) SLA requirement:

1-% averaging in 60-s window Best at 60 s
1/10000 Minimum picking (600-s window) Best at 300 s
MTIE [s]

Best at 600 s
1-% averaging in 300-s window
1E-4
1-% averaging in 600-s window
1-% averaging in
1200-s window

Slave clock 16 ppb

1E-5

10 100 1000 10000

Observation interval [s]
Antti Pietiläinen / 5.11.2008
 Maximum average frequency error (MAFE)
• Regardless of the exact packet selection method used in calculating packet
MAFE the curves remain within a reasonably small range.
• The packet MAFE and MAFE of the clock coincide at large averaging
windows .
fpMAFE, 1 %, 6-s selection window (minimum picking)
1E-4 pMAFE, 1-% percentile
fpMAFE, 1 %, 60-s selection window
1E-5 A candidate mask
MAFE [relative]

packet MAFE for reaching 15 ppb

with 6000-s avera-
1E-6 ging

1E-7
MAFE of packet clock

1E-8

1E-9
1 10 100 1000 10000
tau [s]

Antti Pietiläinen / 5.11.2008

Conclusions of packet timing metrics
• The introduction of packet TDEV in the form of minTDEV was the long-
awaited breakthrough in correlating packet delay variation with packet clock
performance.
• TDEV and MDEV describe average performance. Therefore, not accurately
usable as limit values when occasionally protruding packet delay variations
determine the boundary values of the performance.
• MTIE has been used traditionally also for defining frequency accuracy limits.
However, in this use gives usually worse estimation than actual performance,
incorporating thus unnecessary and variable “safety margin”. Further, it is
difficult to match packet MTIE with packet clock MTIE.
• MATIE and MAFE seem to avoid these issues and the first analysis
considering them as performance estimators and as limiting values have been
promising.
• More analysis with various TIE data and reference slave clocks are still
needed.

Antti Pietiläinen / 5.11.2008

Backup – formulas as pictures
3 TDEV
p. 7. MDEV 

p. 13.
n  k 1
1
MATIEn 0   max
1 k  N  2 n1 n
 x
i k
in  xi  , n = 1, 2, ..., integer part (N/2)

2
N 3n 1 n  j 1 
TDEVn 0  
 N
  xi  2 n  2 xi  n  xi  , n = 1, 2, ..., integer part  
1
6n 2  N  3n  1
 
   3
j 1 i j

 
MTIE(n 0 )  max  max xi  min xi  , n  1,2,... , N  1
1 k  N n k i  k n k i  k n 

n0 is the observation window length, n is the number of samples in the window, 0
is the sample interval, N is the number of samples in the data set. Index variable i is
incremented to scan across the window and k or j is incremented for sliding the
window.

Antti Pietiläinen / 5.11.2008