1. On Modeling Low-Power Wireless Protocols Based On Synchronous Packet Transmissions Marco Zimmerling, Federico Ferrari, Luca Mottola *, Lothar Thiele ETH Zurich * Politecnico di Milano and SICS Swedish ICT TexPoint fonts used in EMF. Read the TexPoint manual before you delete this box.: AAAAAAA

2. Accurate mathematical models of low-power wireless protocols are important »Understanding (e.g., trade-offs, parameters) »Verification »System design (e.g., node placement, power sources) »Prediction and adaptation at runtime But traditional link-based multi-hop protocols are intricate and difficult to model (e.g., ZigBee) Motivation

3. Highly dynamic network topology due to: »Volatile low-power wireless links »Node mobility, failures Traditional protocols maintain substantial network state that governs their operation (e.g., link qualities) »Distributed across nodes »Concurrent, uncontrolled updates Motivation

4. Highly dynamic network topology due to: »Volatile low-power wireless links »Node mobility, failures Traditional protocols maintain substantial network state that governs their operation (e.g., link qualities) »Distributed across nodes »Concurrent, uncontrolled updates Motivation Modeling often stops at the link layer, where interactions are single-hop  insufficient!

5. Example: End-to-End Delivery Devise a model of the packet delivery rate (PDR) observed at the sink PDR = f(  )

6. Example: End-to-End Delivery A B A: B: = network state Media Access Radio Driver wake-up times wake-up times time PDR = f(, …) wake-up times wake-up times one-hop

7. Example: End-to-End Delivery A B = network state Tree Routing Media Access Radio Driver link qualities link qualities wake-up times wake-up times 0.9 0.7 PDR = f(,, …) wake-up times wake-up times link qualities link qualities one-hop one-hop/ end-to-end

8. Example: End-to-End Delivery A B = network state Tree Routing Media Access Radio Driver link qualities link qualities wake-up times wake-up times PDR = f(,, …) wake-up times wake-up times link qualities link qualities one-hop one-hop/ end-to-end

9. Example: End-to-End Delivery A B = network state Application Reliable Transport Tree Routing Media Access Radio Driver queue sizes queue sizes link qualities link qualities wake-up times wake-up times PDR = f(,,, …) wake-up times wake-up times link qualities link qualities queue sizes queue sizes one-hop end-to-end one-hop/ end-to-end

10. Example: End-to-End Delivery A B = network state Application Reliable Transport Tree Routing Media Access Radio Driver queue sizes queue sizes link qualities link qualities wake-up times wake-up times PDR = f(,,, …) wake-up times wake-up times link qualities link qualities queue sizes queue sizes one-hop end-to-end one-hop/ end-to-end Continuously changing

11. Example: End-to-End Delivery A B = network state Application Reliable Transport Tree Routing Media Access Radio Driver queue sizes queue sizes link qualities link qualities wake-up times wake-up times 0.8 0.95 PDR = f(,,, …) wake-up times wake-up times link qualities link qualities queue sizes queue sizes one-hop end-to-end one-hop/ end-to-end Continuously changing

12. Example: End-to-End Delivery A B = network state one-hop end-to-end one-hop/ end-to-end Application Reliable Transport Tree Routing Media Access Radio Driver queue sizes queue sizes link qualities link qualities wake-up times wake-up times 0.8 0.95 Protocols abstract the unreliable wireless channel as point-to-point links  major reason for complexity Continuously changing

13. Multiple nodes send simultaneously to the same receiver, as opposed to pairwise link-based transmissions (LT) Synchronous Transmissions (ST) R R

14. Enabled by IEEE 802.15.4 physical-layer phenomena:  Power (and delay) capture  different/identical packets  Constructive baseband interference  identical packets Several recent protocols exploit ST for various purposes: »A-MAC [SenSys 10]: contention resolution »Glossy [IPSN 11]: network flooding »Splash [NSDI 13]: code updates »CAOS [SenSys 13]: all-to-all communication Synchronous Transmissions (ST)

15. Enabled by IEEE 802.15.4 physical-layer phenomena:  Power (and delay) capture  different/identical packets  Constructive baseband interference  identical packets Several recent protocols exploit ST for various purposes: »A-MAC [SenSys 10]: contention resolution »Glossy [IPSN 11]: network flooding »Splash [NSDI 13]: code updates »CAOS [SenSys 13]: all-to-all communication Synchronous Transmissions (ST) ST enable efficient and reliable multi-hop protocols with very little network state

16. Open Question Do ST also enable simple yet accurate modeling?

17. Modeling Trade-Off Space Accuracy of Models Modeling Scope Tree Routing Radio Driver High Low High

18. Modeling Trade-Off Space Park et al. [IPSN 10] Buettner et al. [SenSys 06] Zimmerling et al. [IPSN 12] Langendoen & Meier [TOSN 10] Polastre et al. [SenSys 04] Challen et al. [MobiSys 10] Tree Routing Radio Driver 2-7 % error Accuracy of Models Modeling Scope High Low High Link layer

19. Modeling Trade-Off Space Park et al. [IPSN 10] Buettner et al. [SenSys 06] Zimmerling et al. [IPSN 12] Langendoen & Meier [TOSN 10] Polastre et al. [SenSys 04] Challen et al. [MobiSys 10] Bruneo et al. [PE 12] Gelenbe et al. [TOSN 07] Guenther et al. [UKPEW 12] Tree Routing Radio Driver 2-7 % error Accuracy of Models Modeling Scope High Low High Link layer Full stack

20. Modeling Trade-Off Space Park et al. [IPSN 10] Buettner et al. [SenSys 06] Zimmerling et al. [IPSN 12] Langendoen & Meier [TOSN 10] Polastre et al. [SenSys 04] Challen et al. [MobiSys 10] Our work Tree Routing Radio Driver 2-7 % error0.25 % error Bruneo et al. [PE 12] Gelenbe et al. [TOSN 07] Guenther et al. [UKPEW 12] Accuracy of Models Modeling Scope High Low High Link layer Full stack

21. Validity of the Bernoulli assumption to ST vs LT Modeling an ST-based multi-hop protocol Model validation through real-world experiments Outline

22. Validity of the Bernoulli assumption to ST vs LT Modeling an ST-based multi-hop protocol Model validation through real-world experiments Outline

23. Intended receiver of a sequence of packets observes a sequence of i.i.d. Bernoulli trials »Success (packet received) with probability p »Failure (packet lost) with probability 1 – p Empirical studies showed that this assumption is not always valid to LT (esp. when packets are close in time) Bernoulli Assumption SR S S Receptions/losses are independent over time

24. Intended receiver of a sequence of packets observes a sequence of i.i.d. Bernoulli trials »Success (packet received) with probability p »Failure (packet lost) with probability 1 – p Empirical studies have shown that this assumption is not always valid to LT Bernoulli Assumption SR S S What about ST? Receptions/losses are independent over time

25. Glossy [IPSN 11] leverages ST for efficient and reliable network flooding Methodology

26. Glossy [IPSN 11] leverages ST for efficient and reliable network flooding Methodology

27. Glossy [IPSN 11] leverages ST for efficient and reliable network flooding Methodology

28. Glossy [IPSN 11] leverages ST for efficient and reliable network flooding Methodology

29. Glossy [IPSN 11] leverages ST for efficient and reliable network flooding Methodology

30. 139-node testbed at NUS 20-byte packets at 20 ms IPI

31. Methodology 139-node testbed at NUS 20-byte packets at 20 ms IPI ST-Type: how protocols perceive Glossy »70 equally distributed nodes »50,000 Glossy floods each »Remaining 138 nodes log

32. Methodology 139-node testbed at NUS 20-byte packets at 20 ms IPI ST-Type: how protocols perceive Glossy »70 equally distributed nodes »50,000 Glossy floods each »Remaining 138 nodes log LT-Type: how protocols perceive LT »All 139 nodes »50,000 broadcasts each »Nodes within radio range log

33. Methodology 139-node testbed at NUS 20-byte packets at 20 ms IPI ST-Type: how protocols perceive Glossy »70 equally distributed nodes »50,000 Glossy floods each »Remaining 138 nodes log LT-Type: how protocols perceive LT »All 139 nodes »50,000 broadcasts each »Nodes within radio range log Tx power: 0 dBm (max) and -15 dBm (lowest possible)

34. Methodology 139-node testbed at NUS 20-byte packets at 20 ms IPI ST-Type: how protocols perceive Glossy »70 equally distributed nodes »50,000 Glossy floods each »Remaining 138 nodes log LT-Type: how protocols perceive LT »All 139 nodes »50,000 broadcasts each »Nodes within radio range log Tx power: 0 dBm (max) and -15 dBm (lowest possible) Collected >1,200,000,000 packet reception events

35. 1)Represent every trace as a discrete-time binary time series {x i } n of length n = 50,000: 010011100111110011101010… 2)Keep only weakly stationary time series, using two empirical tests for non-constant mean/variance 3)Check the validity of the Bernoulli assumption, using a statistical test based on the sample autocorrelation Trace Analysis

36. Bernoulli Results 1010111110111101100110111111101111101110110001111111001111011101011111000111110100111101011001100110 200 1000

37. Bernoulli Results 0.5 0.1 0.03 200 1000

38. Bernoulli Results 40.0 14.5 2.6 0.5 0.1 0.03 200 1000

39. Bernoulli Results 8.7 34.6 4.0 14.1 0.3 2.5 0.5 40.0 0.1 14.5 0.03 2.6 200 1000

40. Bernoulli Results 8.7 34.6 4.0 14.1 0.3 2.5 0.5 40.0 0.1 14.5 0.03 2.6 The Bernoulli assumption is highly valid to ST and significantly more valid to ST than to LT

41. Validity of the Bernoulli assumption to ST vs LT Modeling an ST-based multi-hop protocol Model validation through real-world experiments Outline

42. Low-Power Wireless Bus (LWB) LWB [SenSys 12] uses only ST for communication »Turns a multi-hop network into a “virtual” single-hop network, similar to a shared bus Centralized scheduling »A controller node orchestrates all communication LWB controller multi-hop networkGlossy floodsshared bus

43. Low-Power Wireless Bus (LWB) LWB [SenSys 12] uses only ST for communication »Turns a multi-hop network into a “virtual” single-hop network, similar to a shared bus Centralized scheduling »A controller node orchestrates all communication LWB controller multi-hop networkGlossy floodsshared bus Does the validity of the Bernoulli assumption to ST help devise a simple yet accurate energy model of LWB?

44. A single event − the reception of schedules from the controller − drives the operation of all LWB nodes Schedule-Driven Operation Communication rounds schedule data …

45. Finite-State Machine = schedule received = schedule missed

46. Finite-State Machine = schedule received = schedule missed The Bernoulli assumption allows us to characterize schedule receptions through a single parameter p

47. Discrete-Time Markov Chain p 1-p p p p p pp p p p p = schedule received, with probability p = schedule missed, with probability 1-p p

48. 0 0.2 0.4 0.6 0.8 1 0 0.5 1 Probability of receiving a schedule p Stationary distribution Stationary distribution of the DTMC gives the frequency of visits to each state in the long run for a given p Combined with the well-defined cost of each state, we get the long-term expected energy cost of a LWB node Energy Model

49. Validity of the Bernoulli assumption to ST vs LT Modeling an ST-based multi-hop protocol Model validation through real-world experiments Outline

50. 30-node testbed at ETH Zurich 15-byte payload at 6 sec IPI Tx power: o dBm (max) Estimate several model parameters at runtime, such as: Measure energy consumption using established software- based methods Model Validation p = number of received schedules number of expected schedules

51. Energy Validation Results reality

52. Energy Validation Results reality

53. Energy Validation Results average model error: 0.25% reality artificial

54. Modeling Trade-Off Space Tree Routing Radio Driver Application LWB Glossy Radio Driver Park et al. [IPSN 10] Buettner et al. [SenSys 06] Zimmerling et al. [IPSN 12] Langendoen & Meier [TOSN 10] Polastre et al. [SenSys 04] Challen et al. [MobiSys 10] Our work 2-7 % error0.25 % error Bruneo et al. [PE 12] Gelenbe et al. [TOSN 07] Guenther et al. [UKPEW 12] Accuracy of Models Modeling Scope High Low High Link layer Full stack

55. The Bernoulli assumption is highly valid to ST but often illegitimate to LT ST enable simple yet accurate modeling of a complete state-of-the-art low-power wireless networking stack Conclusions

56. 0 200 400 600 800 1000 Time (seconds) 0.2 0.4 0.6 0.8 1 Packet reception rate Weakly stationary Non-stationary Linear fit Formal tests (e.g., KPSS, ADF) often fail in practice Empirically declare a trace as non-stationary if »PRR changes by 0.015 or more over the entire trace, or »PRR drops/rises by >0.05 within 40 seconds (2,000 packets) Test for Weak Stationarity

57. TypeTx powerTotalNon-stationaryWeakly stationary ST-Type0 dBm9660479613 ST-Type-15 dBm96602569404 LT-Type0 dBm418914182771 LT-Type-15 dBm17775881189 Trace Statistics

58. 0 5 10 15 20 Lag -0.05 0 0.05 0.1 Sample autocorrelation Trace 1 Trace 2 Bounds Let {x i } n a realization of an i.i.d. sequence {X i } ∞ of random variables with finite variance »For large n, about 95% of the sample autocorrelation values should lie within the confidence bounds ±1.96/n 1/2 »We consider the Bernoulli assumption valid for a given trace if the above holds already at lag 1 Validating Bernoulli