Kyeong Soo Kim, Invited talk, 2016 International Conference on Internet of Things and 5G Mobile Technologies (2016 ICIOT-5GMT), Guangzhou University, Guangzhou, Nov. 27-28, 2016.

1. Energy-Efficient Time Synchronization
Achieving Nanosecond Accuracy
in Wireless Networks
Kyeong Soo (Joseph) Kim
(With S. Lee and E. G. Lim@XJTLU)
Department of Electrical and Electronic Engineering
Xi’an Jiaotong-Liverpool University
2016 ICIOT-5GMT
Guangzhou University
27-28 November, 2016
1 / 49

3. Outline
Introduction
Time and Space in Synchronization
Energy-Efficient Time Synchronization for Asymmetric
Wireless Networks
Simulation Results
Next Steps: Extension to Multi-Hop Time Synchronization
Conclusions
3 / 49

4. Next . . .
Introduction
Time and Space in Synchronization
Energy-Efficient Time Synchronization for Asymmetric
Wireless Networks
Hardware and Logical Clock Models
Effect of Clock Skew on Measurement Time Estimation
Asynchronous Source Clock Frequency Recovery at
Sensor Nodes: One-Way Clock Skew Estimation
Simulation Results
Performance of One-Way Clock Skew Estimation
Performance of Measurement Time Estimation and
Energy Efficiency
Effect of Bundling of Measurement Data
Next Steps: Extension to Multi-Hop Time Synchronization
Conclusions
4 / 49

5. An Asymmetric Wireless Network
HeadeNode
Internet
RemoteeUser
SensoreNodes
5 / 49

6. Head Node
A base station that serves as a
gateway between wired and
wireless networks.
A center for fusion of data
from distributed sensors.
Equipped with a powerful
processor and supplied power
from outlet.
6 / 49

7. Head Node
A base station that serves as a
gateway between wired and
wireless networks.
A center for fusion of data
from distributed sensors.
Equipped with a powerful
processor and supplied power
from outlet.
6 / 49

8. Head Node
A base station that serves as a
gateway between wired and
wireless networks.
A center for fusion of data
from distributed sensors.
Equipped with a powerful
processor and supplied power
from outlet.
6 / 49

9. Sensor Node
Measuring data and/or detect
events with sensors and
connected to a WSN only
through wireless channels.
Limited in processing and
battery-powered.
7 / 49

10. Sensor Node
Measuring data and/or detect
events with sensors and
connected to a WSN only
through wireless channels.
Limited in processing and
battery-powered.
7 / 49

11. Design Goals
Achieving sub-microsecond time synchronization
accuracy
Through propagation delay compensation.
With higher energy efficiency at battery-powered
sensor nodes
Minimize the number of packet transmissions and the
amount of computation at sensor nodes.
8 / 49

12. Next . . .
Introduction
Time and Space in Synchronization
Energy-Efficient Time Synchronization for Asymmetric
Wireless Networks
Hardware and Logical Clock Models
Effect of Clock Skew on Measurement Time Estimation
Asynchronous Source Clock Frequency Recovery at
Sensor Nodes: One-Way Clock Skew Estimation
Simulation Results
Performance of One-Way Clock Skew Estimation
Performance of Measurement Time Estimation and
Energy Efficiency
Effect of Bundling of Measurement Data
Next Steps: Extension to Multi-Hop Time Synchronization
Conclusions
9 / 49

13. Effects of Time and Space
The effects of time and space are so closely related that
they cannot be easily separated from each other as in the
following examples:
Synchronization and localization accuracies.
In time-based localization.
e.g. Time of arrival (TOA).
Clock offset and propagation delay.
In one-way synchronization.
e.g. Flooding time synchronization protocol (FTSP).
10 / 49

14. Synchronization and Localization Accuracies
Accuracies
1 ms ↔ 300 km
1 µs ↔ 300 m
1 ns ↔ 30 cm
1 ps ↔ 0.3 mm
Time-based localization schemes
Time of arrival (TOA)
Time difference of arrival (TDOA)
A special variation of TDOA with virtual anchors does
not require synchronization among devices.
⇒ See the next slide.
11 / 49

15. TDOA with Virtual Anchors 1
Anchor
Agent
Virtual
Anchors
1
E. Leitinger et al., IEEE J. Sel. Areas Commun., vol. 33, no. 11, pp.
2313–2328, Nov. 2015.
12 / 49

16. Clock Offset and Propagation Delay
Can the receiver distinguish between the following two
cases if θ = d?
Packet with
Timestamp T vs. Packet with
Timestamp T
TX
RX
TX
RX
• : Clock offset
• : Propagation delay
Answer is “No”.
Two-way message exchanges needed for delay
compensation.
13 / 49

17. Next . . .
Introduction
Time and Space in Synchronization
Energy-Efficient Time Synchronization for Asymmetric
Wireless Networks
Hardware and Logical Clock Models
Effect of Clock Skew on Measurement Time Estimation
Asynchronous Source Clock Frequency Recovery at
Sensor Nodes: One-Way Clock Skew Estimation
Simulation Results
Performance of One-Way Clock Skew Estimation
Performance of Measurement Time Estimation and
Energy Efficiency
Effect of Bundling of Measurement Data
Next Steps: Extension to Multi-Hop Time Synchronization
Conclusions
14 / 49

18. Conventional Two-Way Message Exchanges I
Master
sHead Node)
Slave
sSensor Node)Measurement
Interval of Time Sync. si.e., 2-Way Message Exchange)
……
Report
Request
Response
Report
Measurement
T1
T2
T4
T3
Sensor nodes transmit “Request” messages for
synchronization.
In addition to measurement data packets.
15 / 49

19. Conventional Two-Way Message Exchanges
II
The sensor node can estimate its clock offset w.r.t. the
head node and synchronize its clock to that of the
head node:
Clock offset: ˆθ =
(T2 − T1) − (T4 − T3)
2
.
Propagation delay: ˆd =
(T2 − T1) + (T4 − T3)
2
.
16 / 49

20. Reverse Two-Way Message Exchanges I
Master
sHead Node)
Slave
sSensor Node)
Beacon/
Request
sMeasurement)
Report/
Response
T1 T4
T3T2
d
tm
Sensor nodes do not transmit any other messages
except “Request/Response” messages.
If there are no measurement data, sensor nodes just
receive messages.
17 / 49

21. Reverse Two-Way Message Exchanges II
The head node can estimate the clock offset of the
sensor node, but the sensor node cannot.
As a result, the information of all sensor node clocks
is centrally managed at the head node.
“Response” (synchronization) and “Report”
(measurement data) messages can be combined to
save the number of message transmissions from the
sensor node.
Optionally measurement data and corresponding
timestamps can be bundled together in a
“Report/Response” message when there are no strict
timing requirements.
18 / 49

22. Next . . .
Introduction
Time and Space in Synchronization
Energy-Efficient Time Synchronization for Asymmetric
Wireless Networks
Hardware and Logical Clock Models
Effect of Clock Skew on Measurement Time Estimation
Asynchronous Source Clock Frequency Recovery at
Sensor Nodes: One-Way Clock Skew Estimation
Simulation Results
Next Steps: Extension to Multi-Hop Time Synchronization
19 / 49

23. Hardware Clock Model
Time Ti of the hardware clock of the ith sensor node at the
reference time t is modeled as a first-order affine function:
Ti(t) = (1 + i)t + θi,
where
(1 + i) ∈ R+: Clock frequency ratio.2
θi ∈ R: Clock offset.
2
i is called a clock skew in the literature.
20 / 49

24. Logical Clock Model
Time Ti of the logical clock of the ith sensor node at
hardware clock time Ti(t) is modeled as a piecewise linear
function: For tk<t≤tk+1 (k=0, 1, . . .),
Ti Ti(t) = Ti Ti(tk) +
Ti(t) − Ti(tk)
1 + ˆi,k
− ˆθi,k,
where
tk: Reference time when a kth synchronization occurs.
ˆi,k: Estimated clock skew from the kth
synchronization.
ˆθi,k: Estimated clock offset from the kth
synchronization.
21 / 49

25. Next . . .
Introduction
Time and Space in Synchronization
Energy-Efficient Time Synchronization for Asymmetric
Wireless Networks
Hardware and Logical Clock Models
Effect of Clock Skew on Measurement Time Estimation
Asynchronous Source Clock Frequency Recovery at
Sensor Nodes: One-Way Clock Skew Estimation
Simulation Results
Next Steps: Extension to Multi-Hop Time Synchronization
22 / 49

26. Measurement Time Estimation Error:
Conventional Two-Way Message Exchanges
Master
sHead Node)
Slave
sSensor Node)
Measurement
Request
Response
Report
s1
s2≈s3
s4
d
tm
When Tm d,
∆ˆtConv.
m ∼ Tm × ∆ˆi,
where ∆ˆi is the clock skew estimation error.
23 / 49

27. Measurement Time Estimation Error:
Reverse Two-Way Message Exchanges
Master
sHead Node)
Slave
sSensor Node)
Beacon/
Request
sMeasurement)
Report/
Response
T1 T4
T3T2
d
tm
When Tm d,
∆ˆtRev.
m ∼
Tm
2
× ∆ˆi.
24 / 49

28. Next . . .
Introduction
Time and Space in Synchronization
Energy-Efficient Time Synchronization for Asymmetric
Wireless Networks
Hardware and Logical Clock Models
Effect of Clock Skew on Measurement Time Estimation
Asynchronous Source Clock Frequency Recovery at
Sensor Nodes: One-Way Clock Skew Estimation
Simulation Results
Next Steps: Extension to Multi-Hop Time Synchronization
25 / 49

29. Message Departure and Arrival Times
Let td(k) (k=0, 1, . . .) be the reference time for the kth
message’s departure from the head node.
td(k) also denotes the value of the timestamp carried
by the kth message.
Then the arrival time of the kth message with respect
to the ith sensor node’s hardware clock is given by
ta,i(k) = Ti (td(k)) + d(k) = (1 + i)td(k) + θi + d(k),
where
d(k): One-way propagation delay in terms of the ith
sensor node’s hardware clock.
26 / 49

30. Joint Maximum Likelihood Estimators
For a white Gaussian delay d(k) with known mean d and
variance σ2
,
ˆθML
i (k) =
t2
d
· ta,i − td · tdta,i
t2
d
− td
2
− d,
ˆRML
i (k) =
tdta,i − td · ta,i
t2
d
− td
2
,
where
x k
j=0
x(j)
k
,
xy k
j=0
x(j)y(j)
k
.
27 / 49

31. Regression through The Origin (RTO) Model
The problem of asynchronous source clock frequency
recovery (SCFR) can be formulated as a linear RTO model
as follows: For k = 1, 2, . . .,
˜ta,i(k) = (1 + i)˜td(k) + ˜d(k),
where
˜ta,i(k) ta,i(k)−ta,i(0),
˜td(k) td(k)−td(0),
˜d(k) d(k)−d(0).
28 / 49

32. Cumulative Ratio (CR) Estimator
ˆRCR
i (k) =
˜ta,i(k)
˜td(k)
= Ri +
˜d(k)
˜ts(k)
,
where
Ri: Ratio of the ith sensor node hardware clock
frequency to that of the reference clock (i.e., 1+ i).
29 / 49

33. Next . . .
Introduction
Time and Space in Synchronization
Energy-Efficient Time Synchronization for Asymmetric
Wireless Networks
Hardware and Logical Clock Models
Effect of Clock Skew on Measurement Time Estimation
Asynchronous Source Clock Frequency Recovery at
Sensor Nodes: One-Way Clock Skew Estimation
Simulation Results
Performance of One-Way Clock Skew Estimation
Performance of Measurement Time Estimation and
Energy Efficiency
Effect of Bundling of Measurement Data
Next Steps: Extension to Multi-Hop Time Synchronization
Conclusions
30 / 49

34. Next . . .
Introduction
Time and Space in Synchronization
Energy-Efficient Time Synchronization for Asymmetric
Wireless Networks
Simulation Results
Performance of One-Way Clock Skew Estimation
Performance of Measurement Time Estimation and
Energy Efficiency
Effect of Bundling of Measurement Data
Next Steps: Extension to Multi-Hop Time Synchronization
31 / 49

35. Estimated Clock Skews with Gaussian Delays: σ=1 ns
5 10 15 20 25 30 35 40 45 50
Number of Messages
10−21
10−20
10−19
10−18
10−17
10−16
10−15
MSE
RLS
CR
Joint MLE
GMLLE (Two-Way)
LB for CR
CRLB for Joint MLE
LB for GMLLE
32 / 49

36. Estimated Clock Skews with Gaussian Delays: σ=1 µs
5 10 15 20 25 30 35 40 45 50
Number of Messages
10−15
10−14
10−13
10−12
10−11
10−10
10−9
MSE
RLS
CR
Joint MLE
GMLLE (Two-Way)
LB for CR
CRLB for Joint MLE
LB for GMLLE
33 / 49

37. Estimated Clock Skews with AR(1) Delays3
: σ=1 µs
5 10 15 20 25 30 35 40 45 50
Number of Messages
10−14
10−13
10−12
10−11
10−10
MSE
RLS
CR
Joint MLE
GMLLE (Two-Way)
3
ρ = 0.6.
34 / 49

38. Estimated Clock Skews with AR(1) Delays: σ=1 ms
5 10 15 20 25 30 35 40 45 50
Number of Messages
10−8
10−7
10−6
10−5
10−4
MSE
RLS
CR
Joint MLE
GMLLE (Two-Way)
35 / 49

39. Next . . .
Introduction
Time and Space in Synchronization
Energy-Efficient Time Synchronization for Asymmetric
Wireless Networks
Simulation Results
Performance of One-Way Clock Skew Estimation
Performance of Measurement Time Estimation and
Energy Efficiency
Effect of Bundling of Measurement Data
Next Steps: Extension to Multi-Hop Time Synchronization
36 / 49

40. Estimated Frequency Ratio (Sensor Node) and
Measurement Time (Head Node): SI=100 s
-4E-11
-2E-11
0E+00
2E-11
4E-11
FrequencyDifference[ppm]
Proposed (w/ CR)
Two-Way (w/ GMLLE)
0 500 1000 1500 2000 2500 3000 3500
Time [s]
-1E-02
-8E-03
-6E-03
-4E-03
-2E-03
0E+00
2E-03
4E-03
MeasurementTimeError[s]
Proposed (w/ CR)
Two-Way (w/ GMLLE)
Two-Way
37 / 49

41. Estimated Frequency Ratio (Sensor Node) and
Measurement Time (Head Node): SI=1 s
-4E-11
-2E-11
0E+00
2E-11
4E-11
FrequencyDifference[ppm]
Proposed (w/ CR)
Two-Way (w/ GMLLE)
0 500 1000 1500 2000 2500 3000 3500
Time [s]
-1E-04
-8E-05
-6E-05
-4E-05
-2E-05
0E+00
2E-05
4E-05
MeasurementTimeError[s]
Proposed (w/ CR)
Two-Way (w/ GMLLE)
Two-Way
38 / 49

42. Estimated Frequency Ratio (Sensor Node) and
Measurement Time (Head Node): SI=1 ms
-4E-11
-2E-11
0E+00
2E-11
4E-11
FrequencyDifference[ppm]
Proposed (w/ CR)
Two-Way (w/ GMLLE)
0 500 1000 1500 2000 2500 3000 3500
Time [s]
-1E-06
-8E-07
-6E-07
-4E-07
-2E-07
0E+00
2E-07
4E-07
MeasurementTimeError[s]
Proposed (w/ CR)
Two-Way (w/ GMLLE)
Two-Way
39 / 49

43. Effect of SI on Time Synchronization and
Energy Consumption4
Synchronization Skew Estimation Measurement Time
NTX NRX
Scheme MSE Estimation MSE
Proposed
SI=100 s 8.8811E-25 5.8990E-19 100 36
SI=1 s 9.1748E-25 5.4210E-19 100 3600
SI=10 ms 1.0887E-24 4.7684E-19 100 360100
Two-Way with GMLLE
SI=100 s 1.9021E-24 4.7784E-19 136 36
SI=1 s 1.7034E-24 6.1452E-19 3700 3600
SI=10 ms 9.0992E-25 4.0485E-19 360100 360000
Two-Way
SI=100 s
N/A
3.4900E-05 136 36
SI=1 s 3.4564E-09 3700 3600
SI=10 ms 3.3638E-13 360100 360000
4
Estimations are for the samples taken after 360 s (i.e., one tenth of
the observation period) to avoid the effect of a transient period.
40 / 49

44. Next . . .
Introduction
Time and Space in Synchronization
Energy-Efficient Time Synchronization for Asymmetric
Wireless Networks
Simulation Results
Performance of One-Way Clock Skew Estimation
Performance of Measurement Time Estimation and
Energy Efficiency
Effect of Bundling of Measurement Data
Next Steps: Extension to Multi-Hop Time Synchronization
41 / 49

45. Effect of Bundling on Measurement Time Estimation5
0 500 1000 1500 2000 2500 3000 3500
Time [s]
-2.0E-09
-1.0E-09
0.0E+00
1.0E-09
2.0E-09
MeasurementTimeError[s]
NBM=1
NBM=2
NBM=5
NBM=10
5
SI = 1 s.
42 / 49

46. Effect of Bundling on Time Synchronization and
Energy Consumption
Synchronization Scheme
Measurement Time
NTX NRX
Estimation MSE
Proposed
NBM = 1 5.4210E-19 100 3600
NBM = 2 5.1116E-19 50 3600
NBM = 5 3.7504E-19 20 3600
NBM = 10 2.6468E-19 10 3600
In interpreting the results, the following should be
taken into account:
The bundling increases the length of message
payload.
The increased message payload also can affect the
frame errors and the number of retransmissions.
43 / 49

47. Next . . .
Introduction
Time and Space in Synchronization
Energy-Efficient Time Synchronization for Asymmetric
Wireless Networks
Hardware and Logical Clock Models
Effect of Clock Skew on Measurement Time Estimation
Asynchronous Source Clock Frequency Recovery at
Sensor Nodes: One-Way Clock Skew Estimation
Simulation Results
Performance of One-Way Clock Skew Estimation
Performance of Measurement Time Estimation and
Energy Efficiency
Effect of Bundling of Measurement Data
Next Steps: Extension to Multi-Hop Time Synchronization
Conclusions
44 / 49

48. Multi-Hop Extension through Gateways
45 / 49

49. Challenges and Opportunities
Tradeoff between time-translating and
packet-relaying gateways..
The multi-hop extension should be implemented
together with a routing protocol.
As in LEACH protocol6 and its many variations, the
energy efficiency is also critical in the formation of a
hierarchy and the selection of cluster heads (i.e., the
gateway nodes in the multi-hop extension of the
proposed scheme).
6
W. R. Heinzelman et al., Proc. HICSS’00, Jan. 2000, pp. 1–10.
46 / 49

50. Challenges and Opportunities
Tradeoff between time-translating and
packet-relaying gateways..
The multi-hop extension should be implemented
together with a routing protocol.
As in LEACH protocol6 and its many variations, the
energy efficiency is also critical in the formation of a
hierarchy and the selection of cluster heads (i.e., the
gateway nodes in the multi-hop extension of the
proposed scheme).
6
W. R. Heinzelman et al., Proc. HICSS’00, Jan. 2000, pp. 1–10.
46 / 49

51. Challenges and Opportunities
Tradeoff between time-translating and
packet-relaying gateways..
The multi-hop extension should be implemented
together with a routing protocol.
As in LEACH protocol6 and its many variations, the
energy efficiency is also critical in the formation of a
hierarchy and the selection of cluster heads (i.e., the
gateway nodes in the multi-hop extension of the
proposed scheme).
6
W. R. Heinzelman et al., Proc. HICSS’00, Jan. 2000, pp. 1–10.
46 / 49

52. Preliminary Results
47 / 49

53. Next . . .
Introduction
Time and Space in Synchronization
Energy-Efficient Time Synchronization for Asymmetric
Wireless Networks
Hardware and Logical Clock Models
Effect of Clock Skew on Measurement Time Estimation
Asynchronous Source Clock Frequency Recovery at
Sensor Nodes: One-Way Clock Skew Estimation
Simulation Results
Performance of One-Way Clock Skew Estimation
Performance of Measurement Time Estimation and
Energy Efficiency
Effect of Bundling of Measurement Data
Next Steps: Extension to Multi-Hop Time Synchronization
Conclusions
48 / 49

54. Conclusions
Propose an energy-efficient time synchronization
scheme for asymmetric wireless networks achieving
sub-microsecond time synchronization accuracy.
Also, discuss the optional bundling of measurement
data in a “Report/Response” message.
Topics for further study include
Extension to multi-hop synchronization through
packet-relaying or time-translating gateway nodes;
Energy-delay tradeoff and the effect of frame errors
and retransmissions in bundling of measurement
data.
49 / 49

55. Conclusions
Propose an energy-efficient time synchronization
scheme for asymmetric wireless networks achieving
sub-microsecond time synchronization accuracy.
Also, discuss the optional bundling of measurement
data in a “Report/Response” message.
Topics for further study include
Extension to multi-hop synchronization through
packet-relaying or time-translating gateway nodes;
Energy-delay tradeoff and the effect of frame errors
and retransmissions in bundling of measurement
data.
49 / 49

56. Conclusions
Propose an energy-efficient time synchronization
scheme for asymmetric wireless networks achieving
sub-microsecond time synchronization accuracy.
Also, discuss the optional bundling of measurement
data in a “Report/Response” message.
Topics for further study include
Extension to multi-hop synchronization through
packet-relaying or time-translating gateway nodes;
Energy-delay tradeoff and the effect of frame errors
and retransmissions in bundling of measurement
data.
49 / 49

57. Conclusions
Propose an energy-efficient time synchronization
scheme for asymmetric wireless networks achieving
sub-microsecond time synchronization accuracy.
Also, discuss the optional bundling of measurement
data in a “Report/Response” message.
Topics for further study include
Extension to multi-hop synchronization through
packet-relaying or time-translating gateway nodes;
Energy-delay tradeoff and the effect of frame errors
and retransmissions in bundling of measurement
data.
49 / 49

58. Conclusions
Propose an energy-efficient time synchronization
scheme for asymmetric wireless networks achieving
sub-microsecond time synchronization accuracy.
Also, discuss the optional bundling of measurement
data in a “Report/Response” message.
Topics for further study include
Extension to multi-hop synchronization through
packet-relaying or time-translating gateway nodes;
Energy-delay tradeoff and the effect of frame errors
and retransmissions in bundling of measurement
data.
49 / 49