No. (Betteridge's Law of Headlines)

Transcription

1

2

3 No. (Betteridge's Law of Headlines)

4 Is There Any Practical Theory? Roger Wattenhofer ETH Zurich Distributed Computing

5 Theory & Practice OSDI Multimedia SenSys HotNets STOC SPAA FOCS PODC ICALP IPSN Ubicomp Mobicom SIGCOMM SODA EC

6 Theory Meets Practice? Distributed HCI Systems Soft Eng Graphics Security AI/ML Algorithms No Yes [Just my personal observation]

7 Sensor Networks

8 Data Gathering Roger Wattenhofer ETH Zurich Distributed Computing

9 [PermaSense]

10 Efficiency and Reliability

11 reliable efficient [Google Trends]

12 reliable efficient [Google Trends]

13 This paper does a great job at a complete cross-layer design spanning the MAC, link, routing, and application layers to achieve very low power and high reliability for data collection. In some sense this is the first paper I'd give someone working on communication in sensor nets, since it nails down how to do it right. [Matt Welsh, Best of CS 263]

14 Dozer [Burri, von Rickenbach, W]

15 Energy Efficiency sink

16

17 Energy Efficiency sink

18 Energy Efficiency duty cycling, wake up e.g. every 10 seconds parent synchronizes children no network wide synchronization mean energy consumption: 0.066mW, 10y battery

19 Reliability sink

20 Reliability nodes send beacons to reconnect orphans collisions are explicitly accepted availability & reliability: 99% to %

21 Dozer Measurements [Burri, von Rickenbach, W, 2007] [tinynode]

22 Where s the Theory? no network wide synchronization

23 Network Synchronization is Hard

24 Network Synchronization

25 Tree Based Protocols FTSP PulseSync [Lenzen, Sommer, W, TON] Synchronization Error FTSP PulseSync Average (t > 2000s) µs 4.44 µs Maximum (t > 2000s) 249 µs 38 µs

26 Error with Distance FTSP PulseSync

27 Neighbor Synchronization

28 Neighbor Synchronization? Bad neighbor sync Tree-based Algorithms e.g. FTSP Neighborhood Algorithms e.g. GTSP

29 Theorem: Neighbor Sync is Somewhat Hard

30 Model: Drift & Jitter clock rate 1 + εε 1 1 εε t message delay bounded errors (worst-case) dd εε dd + εε

31 Reasonable Time Must Behave! no stopping no jumping

32 Example: Neighbor Sync is Hard sync to fastest neighbor message delay = 1 dd + εε

33 Example: Neighbor Sync is Hard sync to fastest neighbor message delay = 1

34 Example: Neighbor Sync is Hard sync to fastest neighbor message delay = 1

35 Example: Neighbor Sync is Hard sync to fastest neighbor message delay = 1

36 Example: Neighbor Sync is Hard sync to fastest neighbor message delay = 1

37 Example: Neighbor Sync is Hard sync to fastest neighbor message delay = 1

38 Example: Neighbor Sync is Hard sync to fastest neighbor message delay = 1 0 dd εε

39 Example: Neighbor Sync is Hard sync to fastest neighbor message delay = 1 0

40 Example: Neighbor Sync is Hard sync to fastest neighbor message delay = 1 0

41 Example: Neighbor Sync is Hard sync to fastest neighbor message delay = 1 0

42 Sync To Fastest Neighbor: Local Skew Can Be Diameter

43 Average of Neighbors: Local Skew Can Be Diameter Squared

44 Better Protocol?

45 Reminder: Drift & Jitter clock rate 1 + εε 1 1 εε t message delay bounded errors (worst-case) mm εε mm + εε

46 Theorem: Neighbor Sync is Somewhat Hard neighbor sync error = log diameter lower bound: difficult proof matching upper bound: not trivial as well [Lenzen, Locher, W, JACM]

47 Speaking of Synchronization Roger Wattenhofer ETH Zurich Distributed Computing

48

49

50 The Capture Effect

51 Constructive Interference Same Data Different Data [König, W]

52 Accurate Synchronization neighbor sync error: µs transmission timing: µs [König, W]

53 RSS Gain

54 Or

55

56 [König, W]

57 [König, W]

58 Playing With Radios: Alarming Roger Wattenhofer ETH Zurich Distributed Computing

59 Just Send Waves

60 Slotos time [Flury, W, IPSN]

61 Surprisingly Reliable False Positives: 0.8% False Negatives: 0.08%

62 Alarming with Packets?

63

64 [König, W]

65 The Capture Effect

66 Protocol Layering

67 Protocol Layering -64 dbm -70 dbm -75 dbm -81 dbm Layer 4 Layer 3 Layer 2 Layer 1 [König, W]

68 Packet in Packet

69 Naïve Injecting

70 After Clock Sync

71 Symbols Descrambled

72 Measurements Nothing Symbols Sync Both [König, W]

73 Speaking of Power Control Roger Wattenhofer ETH Zurich Distributed Computing

74 Offending the Audience

75 Power Control: Theory vs. Practice trivial lacks strategy outdated old accept resubmit to STOC accept accept accept accept out of scope 10 years ago accept accept

76

77 Power control is old e.g. LTE

78 but

79 Lots of theory progress how to schedule & power wireless transmissions in a network [Moscibroda, W] [Goussevskaia, Halldórsson, W] [Kesselheim]

80 Many Variants and Extensions models on top of SINR robustness results different approximation criteria distributed algorithms etc.

81 84

82

83

84 Is the Theory Practical? about 30% more throughput more reliable communication links (but still too much overhead)

85 Dutch Propositions

86 Proposition In sensor systems, theory practice.* *There are exceptions. Unfortunately, practical research does not seem to believe that these exceptions exist.

87 Sensor Network Theory How many lines of pseudo code Can you implement on a sensor node? My advice: invest your research s in... impossibility results and lower bounds!

88 Summary

89 Professor ETH Zurich Embedded Information Systems