FineGrained Network Time Synchronization using Reference Broadcasts

1
Fine-Grained Network Time Synchronization using
Reference Broadcasts

• slides for cs7210 by David Spain
• (http//www.cc.gatech.edu/dspain/cs7210/slides_FI
  NE_GRAINED_NETWORK.ppt

2
Reference-Broadcast Synchronization

• Nodes send reference beacons to their neighbors
  via a physical layer broadcast.
• Allows highly accurate relative timescales to be
  established or global timescales when used in
  conjunction with an external time source.
• Motivating applications system debugging, sensor
  data fusion, etc

3
2. Related Work

• NTP-like protocols commonalities (2nd paragraph)
• Settling times for GPS synchronized clocks may be
  too long or unavailable. (sensor power/cost
  issues too)
• CesiumSpray similar but unable to unify
  multiple domains w/o a common external
  timesource.
• Liao et al similar time bounds, but they
  require guarantees about the underlying network
• Others similar time bounds, but require tight
  coupling of the application with the MAC (add a
  deterministic bit-detector)

4
3. Traditional Synch. Methods

• Typical server pulses clock value. Also, client
  request / server reply ? client gains one-way
  latency estimate.
• The problem of network time synchronization is
  all about NON-DETERMINISM.

5
3.1 Sources of Synch Error

• SEND TIME msg from host to network interface
• ACCESS TIME contention on the transmit channel
  (contention-based MACs, ex Ethernet).
• PROPAGATION TIME transit, includes queuing
  switching delay at routers.
• RECEIVE TIME delay between arrival at network
  interface notifying the host

6
3.1/3.2

• RBS removes NON-DETERMINISM of Send Access
• Send RBS gives relative time msgs have no
  timestamp
• THINK about receiver vs server synchronized
• Access Simultaneous for all receivers! (phys.
  layer broadcast)
• Propagation assume 0 b/c beacons are only heard
  over small area (is this assumption valid?)
• Receive factor this out as mostly deterministic!
  Time to read a bit is constant, read system clock
  in interrupt handler. Remaining non-determinism
  is Gaussian.

7
4. Reference Broadcast Synchronization (single
hop RBS algorithm)

• Phase Offset (4.1)
• Broadcast a packet to n gt 2 receivers
• Receivers record local time of broadcast
• Receivers trade observations
• 3 Factors
• Group dispersion (max phase error among all rcvr
  pairs)
• Non-determinism of receiver
• Clock Skew
• Receiver non-determinism may be statistically
  removed by sending multiple broadcast beacons
  (Recall non-D was Gaussian)

8
4.2 Estimation of Clock Skew

• Clock Skew clocks run at different rates.
• Accuracy difference between an oscillators
  actual expected frequencies.
• Stability the oscillators tendency to stay at
  the same rate (long short term)
• Calculated least-squares linear regression over
  the multiple phase error samples. Frequency
  phase of the local clock vs remote are given by
  slope line intercept. (Fig 4)
• Allows conversion from local clock to remote
  clock times!

9
4.4 Commodity Hardware Implementation 4.5
Kernal Timestamps

• RBS as a user space daemon on IPAQs
• NTP NTP-Offset account for NTPs delayed
  adjustment
• Tested under light heavy network loads
• Light RBS 8 times better than NTP
• Heavy RBS about the same, NTP -30x
• Timestamping during network interrupt yields
  excellent results (1.85usec 1.28usecs vs
  6.29usec 6.45usec)

10
4.6 Post-Facto Synchronization

• Wake-up after an event of interest
• Perform synchronization with neighbor nodes
• Use least squares linear regression to
  extrapolate backwards
• Saves power in a sensor network NTP is
  particularly unsuited to these deep sleep
  situations (Question if sleeping, how would it
  detect the event of interest?)

11
4.7 Summary of Advantages of RBS

• Largest sources of non-deterministic latency
  removed via using broadcast to synchronize
  receivers
• Multiple broadcast samples provide
• Synchronization b/c errors have a nice
  statistical distribution.
• Estimation of clock skew
• Extrapolation of past phase offsets post-facto
  synchronization
• Outliers lost packets handled best fit line
  still works
• Local timescales (absolute available in the next
  section!)

12
5.1 Multihop Clock conversion

• It is just the single hop algorithm extended!
• A two hop example
• Local clock ? Neighbors clock
• Neighbors clock ? Destination clock
• Requires overlap in the broadcast domains

13
5.2 Time Routing in Multihop Networks5.3
Multihop Performance

• Shortest path using Root Mean Squared error
  (scalability issues in larger networks)
• A more localized routing scheme do time
  conversions as the packet is routed
• Avg n-hop path error approx sn0.5 there are no
  comparison results given!

14
5.4 Synchronization to External Timescales

• Add a host to the broadcast routing domain that
  has GPS time
• Other hosts find a multihop conversion route to
  that timeserver. Easy.

15
7. Summary Future Work

• RBS synchronizes a set of receivers
• Remove non-determinism via broadcast
• Nice error distribution allows calculation of
  clock skew post facto synchronization
• Localized global timescales
• Need
• Algorithm to elect which nodes are beacons
• Minimum time to acquire synchronization