Achieving Network Consistency. Octav Chipara

Transcription

1 Achieving Network Consistency Octav Chipara

2 Reminders Homework is postponed until next class if you already turned in your homework, you may resubmit Please send me your peer evaluations 2

3 Next few lectures Start building a wireless stack from the ground up already covered phy properties mac layer today: network consistency next class: Prof. Ted Herman will talk about timesync future lectures network consistency link estimation topology control routing 3

4 Problem formulation Consistency is a foundation for many network protocols routing tree maintenance => next hop has lower cost network configuration => all nodes have the most recent configuration neighborhood maintenance => a node in all its neighbor s lists Goal: when a node updates/generates a new piece of data, this information must be relayed to all other nodes minimize the number of redundant transmissions (i.e., a node should receive a packet only once) scales well with network size and density 4

5 Basic approach - Flooding Upon hearing new data a node rebroadcasts it eventual consistency Challenge: wireless is a broadcast medium broadcast storm 5

6 Basic approach - Flooding Upon hearing new data a node rebroadcasts it eventual consistency Challenge: wireless is a broadcast medium broadcast storm 5

7 Basic approach - Flooding Upon hearing new data a node rebroadcasts it eventual consistency Challenge: wireless is a broadcast medium broadcast storm 5

8 Basic approach - Flooding Upon hearing new data a node rebroadcasts it eventual consistency Challenge: wireless is a broadcast medium broadcast storm 5

9 Basic approach - Flooding Upon hearing new data a node rebroadcasts it eventual consistency Challenge: wireless is a broadcast medium broadcast storm Broadcast storm: every tries to transmit at the same time resulting in numerous collisions 5

10 Mitigating the broadcast storm problem Randomized delays: introduce delays before packet transmissions reduces the likelihood of packet collisions however, it is often hard to determine the optimal delays depends on the the local node density Transmission suppression: some nodes do not need to transmit a node that hears the same data from several neighbors stops transmitting reduces the number of contending nodes however, it may prolong the time to propagate the message 6

11 Randomized delays C A Reduces likelihood of collisions nodes A, B, C, D transmit at different times B Still inefficient node C and D should not transmit node A and B share many neighbors D 7

12 Transmission suppression C A Transmission suppression reduces the number of contenders potential for significant savings B Knowing more information may help you make better decisions e.g., two hop neighborhood info D 8

13 Transmission suppression C A Transmission suppression reduces the number of contenders potential for significant savings B Knowing more information may help you make better decisions e.g., two hop neighborhood info D 8

14 Transmission suppression C A Transmission suppression reduces the number of contenders potential for significant savings B Knowing more information may help you make better decisions e.g., two hop neighborhood info D 8

15 Transmission suppression C A B Suppressing wrong transmissions will increase propagation delays e.g., suppressing A and B stops progress D 9

16 Trickle - algorithm outline Divides the time into intervals, nodes are synchronized a node transmits metadata in each interval In response to a change in metadata a node picks a random time in its current interval t to transmit its data let c be the number of times a node hears a data item if c < threshold, then node transmits the data item else, transmission is suppressed 1

17 Example k=1 c A t A1 t A2 B t B1 t B2 C τ t C1 t C2 time transmission suppressed transmission reception k = 1 11

18 Example counter k=1 c A t A1 t A2 B t B1 t B2 C τ t C1 t C2 time transmission suppressed transmission reception k = 1 11

19 Example counter k=1 c A t A1 t A2 B t B1 t B2 C τ t C1 t C2 time transmission suppressed transmission reception suppression k = 1 threshold 11

20 Example c A t A1 t A2 B 1 t B1 t B2 C τ t C1 t C2 time transmission suppressed transmission reception k = 1 12

21 Example k=1 c A t A1 t A2 B 2 t B1 t B2 C τ t C1 t C2 time transmission suppressed transmission reception IETF 74 8 k = 1 13

22 Example k=1 c A t A1 t A2 B 2 t B1 t B2 C τ t C1 t C2 time transmission suppressed transmission reception k = 1 14

23 Example c A t A1 t A2 B t B1 t B2 C τ t C1 t C2 time transmission suppressed transmission reception k = 1 15

24 Example k=1 c A 1 t A1 t A2 B t B1 t B2 C 1 τ t C1 t C2 time transmission suppressed transmission reception k = 1 16

25 Example k=1 c A 1 t A1 t A2 B t B1 t B2 C 1 τ t C1 t C2 time transmission suppressed transmission reception 12 k = 1 17

26 Example k=1 c A 1 t A1 t A2 B t B1 t B2 C 1 τ t C1 t C2 time transmission suppressed transmission reception k = 1 18

27 Trickle - features Managing protocol overhead it is wasteful to transmit state information when nothing changes insight: upon a change metadata should be transmitted fast after a change rate of transmitting metadata should decrease solution: exponentially decrease the metadata when state is consistent reset the rate of transmitting metadata upon hearing new data Suppression based on number of overhead packets relies on minimal topological information tolerates frequent changes in topology 19

28 Impact of packet losses [Single hop network] The number of rounds scale with O(log(n)) 2

29 Tickle without synchronization Remove the requirement of nodes operating synchronized each node operates independently the intervals are not aligned anymore New problem: short-listening 21

30 Short-listening problem B transmits soon after the start of each interval reduce likelihood for its transmissions to be suppressed 22

31 Solution to the short-listen problem Divide a slot in two parts listen only - nodes only listen during this part of the interval transmit part - nodes transmit randomly within this interval 2 listen transmit 23

32 Simulation results 5 spacing, 6 hops 2 spacing, 4 hops 24

33 How could we improve Trickle? Take advantage of the spatial correlation of packets Differentiate between stable and unstable neighbors 25