Time synchronization in sensor networks

1
Time synchronization in sensor networks

• Presented by Nirav Jasapara
• Graduate Student, USC.
• jasapara_at_isi.edu

2
Table of Contents

• Time Synchronization Introduction.
• Network Time Protocol (NTP).
• Challenges for time-sync in Sensor networks.
• Reference Broadcast System.
• Time-sync protocol for sensor networks.
• Summary.
• Conclusion

3
Introduction

• What is time-synchronization ?
• A method which allows individual entities in
  a group to synchronize their clocks w.r.t each
  other or to some coordinated universal time
  (UTC).
• Why do we need time-synchronization ?
• Agreeing on a common time scale
• Track relative order of events
• Co-ordinate future events
• Order logged events for debugging
• Handling leases

4
How is it done ?

• Wired Domain
• -- NTP (Network Time Protocol)
• - most widely used time-synchronization
  protocol for
• large multi-hop networks.
• - allows construction of a hierarchy of time
  servers
• rooted at sources of external time (eg. UTC)
• Wireless Domain
• -- RBS (Reference Broadcast System)
• -- TPSN (Time-sync protocol for sensor networks)

5
Network Time Protocol

• Used since 1985 for time-synchronization over
  WANs.
• Unix NTP daemon ported to almost every server
  platform available today - from PCs to Crays -
  Unix, Windows, VMS and embedded systems.
• Provides clients with an average accuracy below 5
  ms.
• Stands out by virtue of its scalability,
  self-configuration in large multi-hop networks,
  robustness to failures and ubiquitous deployment.

6
NTP A Brief Overview

• Uses a hierarchical design.
• At the top are stratum 1 clocks, computers with
  source of accurate time (eg. GPS, WWVB).
• Stratum 2 clocks synchronize themselves with
  stratum 1 clocks over the network.
• This process repeats up to stratum 16 which is
  effectively infinity for NTP.
• A typical NTP network would have stratum 1
  somewhere on the internet, stratum 2 clocks near
  a gateway and stratum 3 clocks on a LAN.
• NTP network is semi-self-organizing each node
  requires some manual configuration, once the NTP
  host is set up it can run itself autonomously.

• Figure 13 A schematic design of
• an NTP network

7
Sensor Networks A New Domain

• Wireless Sensor Networks (WSN) consists of large
  numbers of cooperating small-scale nodes, spread
  over a geographical area, each capable of limited
  computation, wireless communication, and sensing.
• Why existing solutions for wired networks (NTP)
  wont work ?
• - Energy constrained.
• - Dynamic topology.
• - No infrastructure support.

8
Reference Broadcast System (RBS)

• Developed by
• Jeremy Elson, Lewis Girod and Deborah Estrin.

9
Reference Broadcast System

• Exploits the broadcast channel available in many
  physical-layer networks.
• -- Works by sending periodic reference beacons
  to a set of receivers using physical layer
  broadcasts.
• -- Broadcasts do not contain a time-stamp.
• -- Broadcasts used as a relative time reference
  to synchronize a set of receivers.
• Scheme is receiver-receiver as opposed to
  sender-receiver (NTP).
• Tries to achieve accuracy by removing the
  senders non-determinism from the system.
• Federate clocks over different broadcast domains
  with little loss of accuracy over multiple hops.

10
Reference Broadcast System

• The enemy of precise network time-sync is
  non-determinism in latency which contributes
  directly to synchronization error. 1
• Sources of Latency include
• Send Time
• Time required to construct the message
• Access Time
• Time required to access the medium
• Propagation Time
• Time taken for beacon to travel from one NIC card
  to the other
• Receive Time
• Time taken to receive the beacon and notify the
  host of its arrival
• Broadcast at the physical layer will arrive at a
  set of receivers with very little variability in
  its delay. This reduces non-determinism due to
  send-time and receive-time.

• Figure 2 1 Critical path analysis for
  traditional protocols.

11
Estimation of Phase Offset

• Sequence of Events
• A transmitter broadcasts a reference packet to
  two receivers (i and j).
• Each receiver records the time that the reference
  was received, according to its local clock.
• The receivers exchange their observations.
• Based on this single broadcast the receivers can
  form a relative timescale.

12
Estimation of Phase Offset

• The receiver error is Gaussian.
• We can increase precision by sending more
  reference beacons.

• A transmitter broadcasts m reference packets.
• Each of the n receivers records the time that the
  reference was observed, according to its local
  clock.
• The receivers exchange their observations.
• Each receiver i can compute its phase offset to
  any other receiver j as the average of the phase
  offsets implied by each pulse received by both
  nodes i and j. That is, given
• n the number of receivers
• m the number of reference broadcasts
• Tr,b rs clock when it received broadcast b

13
Estimation of Clock Skew

• What is Clock Skew ?
• The change in frequency of the oscillator over
  time.
• Why is it necessary ?
• Clock Skew is major source of error.
• Oscillator Characteristics
• Accuracy
• Agreement between the oscillators actual and
  expected frequencies
• Stability
• An oscillators tendency to stay at the same
  frequency over a period
• To calculate clock skew, least squares linear
  regression is performed
• The slope of the best fit line through the phase
  offsets can give the relative clock skew of the
  remote node.

14
Implementation on Berkeley Motes

• First test
• 5 Berkeley motes sending out reference pulses
  periodically at 2 micro sec clock resolution.
• Residual error detected 11.2 micro sec
• Second Test
• Used motes as NIC cards for std. Linux boxes.
• Modified Linux kernel to time-stamp serial port
  interrupts for reduction in phase calculation
  error.
• Tried to test the clock skew measurement
  algorithm by sending a few synchronization
  pulses, then silencing the radio for a minute and
  sending synchronization pulses again.
• Residual error detected 7.6 micro sec

15
Relative Performance of RBS and NTP

• Implemented RBS under same constraints as NTP.
• Ran RBS as a UNIX daemon on a Compaq IPAQ, with
  Familiar Linux running, using UDP datagrams as
  motes.
• Daemon ran completely in user space, using
  standard system calls the same as NTP.
• Three different synchronization schemes tested
• RBS
• NTP
• NTP with offset
• Test application queried the NTP daemon for the
  clocks phase offset and subtracted it from the
  test pulse time.
• Two Different scenarios tested
• Light Network Load
• Little or not background traffic
• Heavy Network Load
• Two additional IPAQs configured as traffic
  generators sending UDP datagrams at 6.5Mbit/sec.

16
Test Results

• RBS performs 8 times better than NTP in the same
  environment.
• The performs of RBS remains almost the same even
  with heavy traffic
• The performance of NTP degrades 32 fold in the
  presence of heavy traffic
• Figure 5 1

17
Application to Post-Facto synchronization

• What is Post-Facto synchronization ?
• Post-facto synchronization strives to conserve
  energy in sensor nodes by not keeping the clocks
  in continuous synchrony. Clocks are synchronized
  only when an event of interest occurs.
• RBS clock skew estimator acts as an effective
  form of post-facto synchronization.
• Once nodes power up from deep sleep and exchange
  updates, exact time of events of interest can be
  calculated using backwards extrapolation of
  clock-skew.

18
Multi-Hop Time Synchronization

• Let the notation Ei(Rj) mean the time of event i
  according to receiver js clock.
• Receivers R1 and R7 observe events at times
  E1(R1) and E7(R7), respectively.
• R4 uses As reference broadcasts to convert clock
  values from R1 to R4. This is used to convert
  E1(R1) to E1(R4).
• R4 similarly uses Bs broadcasts to convert
  E1(R4) to E1(R7).
• The time elapsed between the events is computed
  as E1(R7)-E7(R7)

19
Synchronization with external timescales

• RBS can be extended to include external
  timescales.
• Only one node is required to be as the reference
  external timescale, for e.g. a special node can
  be GPS enabled and can synchronize its own clock
  externally.
• Then other nodes can convert their time to GPS
  times using the conversion algorithms.

20
Scrutiny RBS

• Advantages
• Receiver-Receiver mechanism, hence most sources
  of non-deterministic latency removed giving high
  precision.
• Multiple reference broadcasts reduce
  synchronization offset.
• Allows nodes to construct local timescales which
  is sufficient for applications like sensor
  networks
• Time can be propagated across broadcast domains.
• Limitations
• It requires a network with a physical broadcast
  channel
• It can not be used, for example, in networks that
  employ point-to-point links.

21
Time Synch Protocol for Sensor Networks (TPSN)

• Developed by Saurabh Ganeriwal, Ramkumar
  Rengaswamy and Mani B Srivastava.

22
Time Sync Protocol for Sensor Networks

• Extend NTPs approach for Time Synchronization in
  wireless sensor networks.
• Uses the sender-receiver paradigm of NTP
• Works in two steps
• Sets up Level hierarchy (Level discovery phase)
• Propagates synchronization messages
  (Synchronization Phase)
• Tries is to maintain a unique global timescale,
  at every instant of time, throughout the network.

23
Level Discovery Phase

• Root node assigned level 0
• The root node initiates this phase by
  broadcasting a level-discovery packet
• Any neighbor receiving this packet will set its
  own level to one greater than the one they
  received and send its own level discovery packet.
• This process is continued and eventually every
  node in the network is assigned a particular
  level.
• Once a nodes level has been decided, it ignores
  all such packets in the future

24
Synchronization Phase

• Root node sends out a time-sync packet.
• Pair-wise synchronization is performed over the
  entire network.
• Figure 8
• At time T1, A sends a sync pulse packet to B
  which contains the level number of A and T1.
• Node B receives this packet at T2, where T2 is
  equal to T1 D d.
• Dclock drift between the two nodes
• dpropagation delay
• At time T3, B sends back an acknowledgement
  packet to A with values of T1, T2 and T3.
• A can calculate the phase offset as shown in
  figure 9.

• Figure 8 2 Two Way Message
  Exchange

25
Some Special Provisions

• TPSN has some special provisions for
• when a new node joins a network or when it does
  not get a level discovery packet
• Wait for some time and listen for a level
  discovery packet
• Timeout and Broadcast a Level Request query
• Neighboring nodes send their own level
• Again select the lowest level received i and
  assign yourself i1
• when a node i loses all its neighbors with
  level i1
• After sending synchronization initiation messages
  a fixed number of times, a node times out and
  sends out a Level Request query
• when the root node dies
• Level 1 nodes will timeout on ack packets to
  root.
• Instead of Level Request query, they run an
  election algorithm and select a new root node.
• A brand new Level Discovery phase is started till
  all the nodes get reassigned their new level.

26
Error Analysis in TPSN

• Decomposition of packet delay
• Figure 9 2

27
Implementation on Berkeley Motes

• Tested on IPAQ testbed with no external
  components used.
• Modification to TinyOS
• Generate a lower granularity clock.
• Timestamp packets at the MAC layer.

28
Test Results

• Contribution of uncertainty at the sender towards
  the sync-error is very low (0.65 microsec) thus
  can be ignored.
• Comparison with RBS showed that on an average it
  was twice as accurate as RBS under similar
  conditions.

• Figure 12 2 Statistics of Sync-error

29
Multi-Hop results

• Figure 13 2 Sync-error over Multi-Hop
• It was observed that error does not blow up as
  the hop count increases.
• Some experiments showed high error values but its
  probability is low.
• Better analytical models for multi-top topologies
  are under research.

30
Scrutiny TPSN

• Advantages
• Being a logical extension of NTP to sensor
  networks, supports high level of scalability.
  Error is almost entirely based on hop distance of
  adjacent nodes and independent of total number of
  nodes in the network.
• Can achieve better accuracy by using a min.
  spanning tree algorithm (level discovery phase)
  with an added complexity of O(N).
• Limitations
• Requires MAC layer time-stamping.
• The main drawback is the time taken for
  synchronization phase which increases linearly
  with N (the number of nodes).
• This initial setup might take a long time and
  might have to be done more than once especially
  in dynamic situation.
• Performance in the presence of traffic is not
  investigated.

31
Summary
32
Conclusion

• NTP is the most widely used time synchronization
  protocol over the large WANs and other wired
  networks
• RBS works very well in broadcast networks (eg.
  WSN) giving higher precision as compared to NTP.
• TPSN improves accuracy over RBS by a factor of 2
  but performance in actual sensor network
  environments is yet to be tested.

33
Acknowledgements

• Dr. Wei Ye
• Dr. Jeremy Elson

34
References

• Jeremy Elson, Lewis Girod and Deborah Estrin,
  Fine-Grained Network Time Synchronization using
  Reference Broadcasts, In Proceedings of the Fifth
  Symposium on Operating Systems Design and
  Implementation (OSDI 2002), Boston, MA. December
  2002.
• Saurabh Ganeriwal, Ramkumar Rengaswamy, Mani B
  Srivastava, Timing-sync Protocol for Sensor
  Networks, ACM SenSys'03, November 2003.
• D.L. Mills, Precision synchronization of computer
  network clocks, ACM Computer Communication Review
  24, 2, April 1994.

35
Questions ?

• Thank You.

36
Backup Slides
37
Estimation of Phase Offset

• Figure 4 1

• Figure 4 shows the maximum pair-wise error after
  RBS. The receiver reports errors according to the
  Gaussian distribution.
• Figure 4-a shows the mean group dispersion and
  the standard deviation from the mean.
• Figure 4-b shows a 3D view of the same dataset,
  from 2 to 20 receivers

38
RBS using kernel timestamps.

• Time-stamping at interrupt time, before the
  packet is even transferred from the NIC,
  significantly reduces jitter.
• Kernel Time stamping significantly improved the
  performance over the same setup.
• 1.85 1.28 microseconds from 6.29 6.45
  microseconds.

39
Time Routing in Multi-Hop Networks

• Figure 7 1 A more complex multi-hop
  network topology.