1. Wireless Sensor Networks Clock Synchronization Professor Jack Stankovic University of Virginia

2. Why is clock sync important Spatial-Temporal System Where and when events occur To distinguish multiple events, compute lifetime of event, coordinate sensing, coordinate control actions (for wake-up, for actuators, …), … Signal processing (local clock may be enough) Compute velocity

3. Could we have a WSN without clock sync? –If all operations are logical Or Use GPS! –Power hungry –Expensive per node

4. Outline NTP (Network Time Protocol) – for Internet RBS (Reference Broadcast Sync) TPSN (Time-sync Protocol for Sensor Networks) FTSP (Flooding Time Sync Protocol)

5. What is Clock Sync The clocks of each node in the WSN should read the same time within epsilon and remain that way –Sync clocks –Handle clock drift (up to 40 microsec per second on Mica platform; 7.37 MHz clock) Solution: Re-sync (expensive) Solution: Estimate clock drift and compensate Drift Resync infrequently

6. How bad is 40 microsec/sec? 40 x 60 = 2,400 microsec/min 2,400 x 60 = 144,000 microsec/hr 144,000 x 24 = 3,456,000 microsec/day

7. Goals for Clock Sync in WSN Meet required precision Low communication BW Robust against node and link failures Multi-hop (large scale)

8. NTP Network Time Protocol (NTP) on Internet –Included in OS code –Simple NTP (SNTP) exists for PCs –Runs as background process –Get time from “best” servers Finds those with lowest jitter/latency Depends on wired and fast connections –64 bit timestamps sent (200 picosec resolution) –Use statistical analysis of round trip time to clock servers –Clock servers get time from GPS –Millisec (ms) accuracy (per hop)

9. NTP Not (directly) usable for WSN –Complex – see 50 page document –Continuous cost –Large code size –Expensive in messages/energy –Non-determinism at MAC layer (per hop) if NTP was implemented in WSN can add 100s ms delay

10. Clock Sync Delays Uncertainties of radio message delivery

11. Clock Sync - Delays Send Time – assemble message and issue request to MAC layer –System call overhead –OS may delay the call (how busy is the node) Access Time – waiting for access to the channel Transmission Time – time to transmit; F(speed of radio, length of message)

12. Clock Sync - Delays Propagation Time – time to get from the sender to receiver; less than 1 microsec for ranges up to 300 m. Reception Time – time for receiver to collect message from the airways; same as transmit time Receive Time – time to process the incoming message

13. Ideal Read Clock 5 5 Set Clock All Zero Costs/Delays Propagates 5 5 Real Clock Real Clock

14. What if Deterministic Delays Read Clock 5 5 Set Clock Propagates 5 5 Real Clock Real Clock Minimize these times Make deterministic.1.01.1 5.21 ? ?

15. RBS Reference message is broadcast Receivers record their local time when message is received –Timestamps are ONLY on the receiver side –This eliminates access and send times Nodes exchange their recorded times with each other 29.1 microsec accuracy for 1 hop case No transmitter side non-determinism Not extended to large multi-hop networks

16. RBS 567567 Local Time Send MSG No Clock 6,7 5,7 5,6 ExchangeAdjust Plus 2 Plus 1 OK Propagation Time = 0

17. What can cause the “ideal” case of all three nodes getting the clock message at the “same” time to be false? –Node turns off interrupts Code explicitly turns off interrupts nesC atomic statement –Node turns off radio (sleep mode)

18. TPSN Creates a spanning tree Perform pairwise sync along edges of the tree Must be symmetric links –Recall radio irregularity paper Considers “entire” system –RBS looks at one hop Exchange two sync messages with parent node

19. Synchronization Phase Delta = clock drift P = propagation delay T1 T2 T3 T4 Node A Node B P = ((T2-T1)+(T4-T3))/2 Delta = ((T2-T1)-(T4-T3))/2 Node A corrects its clock by Delta Note: Sender A corrects to clock of receiver B T2=T1+P+Delta (T1)(T1,T2,T3)

20. Example T1= 5.0 T2=5.35T3=6.0 T4=5.75 Node A Node B P = ((T2-T1)+(T4-T3))/2 Delta = ((T2-T1)-(T4-T3))/2 P = ((5.35-5.0)+(5.75-6))/2 Delta = ((.35)-(-.25))/2 P =((.35)+(-.25))/2 Delta = 0.6/2 =.3 P =.1/2=.05 5.3.05 5.7 So A adds.3 to 5.75 to get 6.05 Only need Delta to adjust clocks 5.05

21. TPSN No broadcasting so TPSN is expensive Timestamp the message in the MAC layer multiple times and average those times (improves accuracy over RBS by 2 times) (16.9 microsec accuracy for 1 hop) No clock drift corrections

22. Read Clock in MAC Layer Uncertainties of radio message delivery

23. TPSN Assumes topology does not change –Not robust to failure of a node Node fails – must redo the spanning tree

24. Flooding Time Sync Protocol (FTSP) Implemented on Mica platform ~1 Microsec accuracy MAC-layer timestamp Skew compensation with linear regression (accounts for drift) Periodic flooding – robust to failures and topology changes Handles large scale networks

25. Remove Uncertainties Eliminate Send Uncertainty –Get time in the MAC layer Eliminate Access Time –Get time after the message has access to the channel Eliminate Receive Time –Record local time message received at the MAC layer

26. Remaining (Mostly) Deterministic Times –Transmit –Propagation –Reception

27. Basic Idea When to time stamp the message sender receiver Mica2 uncertainties: Interrupt:5us, 30us(<2%) Encode+decode:110us -112us Byte align: 0us – 365us Propagation: 1us Get timestamp Set timestamp Ready

28. Basic Idea When to time stamp the message –Radio layer, after the second SYN sent out, 6 timestamps in row, take the average and send only 1 timestamp Normalize and then take Average of these timestamps for 6 bytes of data

29. Why take 6 samples? Because of all the uncertainties as described in a previous slide

30. FTSP Root maintains global time for system All others sync to the root Nodes form an ad hoc structure rather than a spanning tree Root broadcasts a timestamp for the transmission time of a certain byte Every receiver time stamps reception of that byte Account for deterministic times Differences are the clock offsets

31. Summary of FTSP So Far (Below) MAC-layer timestamp Correct sender timestamp to account for known delays –Compute final offset error Result: 1.48 microsec accuracy for 1 hop

32. Clock Drift If local clocks all had exact same frequency then no clock drift corrections are needed and no re-syncs needed either MICA2 motes –Crystals used have drifts of up to 40 microseconds per second –Implication: resync every sec to maintain 40 microsec accuracy –Too costly - estimate clock drift and fix!

33. Basic Idea 8 entry linear regression table to estimate clock skew (each entry derived from 1 clock sync protocol execution) Example 15 us offset 25 us offset 37 us offset 44 us offset 54 us offset 65 us offset 77 us offset 85 us offset 5 us over Say a 10 sec Resync period Compensate.5 us Each sec 10 seconds 1 sec

34. Performance Single hop, stop sync after 33 minutes (where sync was run every 30 secs) Example: If clock sync stops at time 33 and error of 10 Microsec is OK; then 7 minute re-sync cycle is OK

35. Multi-Hop Time Sync Create a fixed spanning tree for clock sync dissemination like in TPSN (not robust) Use more ad hoc technique –Assume every node has unique ID –One root at a time

36. Root/Reference Point Step 1 Step 2 8 sync msgs to perform Linear regression How to handle messages from multiple nodes

37. Sync Message Format Timestamp – global time estimate of transmitter when broadcast RootID seqNum – incremented by root when new sync round is initiated (e.g., every 30 seconds) –Other nodes insert largest received seqNum into messages they broadcast

38. Processing Redundant Messages Place only first neighbor message for each sequence number in linear regression table Other redundant messages (with same sequence number) are ignored Once 8 messages are obtained then can perform linear regression –Note: the 8 messages may come from different neighbors

39. Root Election Process Option – base station is root – done, but not fault tolerant Multiple base stations? Any node can serve as root? If no sync message for time T, declare myself to be root May be multiple roots Smallest ID wins; so roots will give up their root status when it eventually gets a sync message from another root with lower ID

40. Performance/Example Multiple-hop, sync every 30 seconds A) at 0:00 all motes were turned on; B) at 0:41 the root with ID1 was switched off; C) from 1:12 until 1:42 randomly selected motes were switched off and back on, one per 30s; D) at 1:47 the motes with odd node IDs were switched off (half of the nodes are removed); E) at 2:02 the motes with odd node IDs were switched back on; F) at 2:13 the second root with ID2 is switched off; ID1 (initial root - elected) ID2

41. Performance (not real data) 6 minutes to elect root, 10 minutes complete sync, 1.2ms per hop

42. Questions How long does it take to clock sync the entire system (of say 1000, or 10,000 nodes)? –60 nodes: 10 minutes to elect leader; 14 minutes for sync –Collisions What happens during the election process?

43. Questions What if some nodes do not have synchronized clocks? –Lost messages –Down nodes (and come back)

44. Summary Very accurate clock sync is possible (in microsec range) Time to perform clock sync is important –Initial delay at system init time may be important Consider a system of 10,000 nodes Multiple base stations issues Solve clock drift –Frequent sync messages –Infrequent sync messages but adjust locally

45. Summary Consider costs –Energy –Congestion –Time Consider accuracy requirement FTSP available as a nesC component