Searchlight Wont You Be My Neighbor Mehedi Bakht
Searchlight: Won't You Be My Neighbor? Mehedi Bakht, Matt Trower, Robin Kravets Department of Computer Science University of Illinois Robin Kravets, University of Illinois
Is anybody out there? 2 Robin Kravets, University of Illinois
Is anybody out there? Registration services Foursquare, Facebook, Google Latitude - centralized, slow, difficult to manage across apps Provides applications with absolute locations 3 Robin Kravets, University of Illinois
Is anybody out there? Direct mobile-to-mobile communication Qual. Comm All. Joyn, Nokia Sensor, Nintendo Street. Pass, Sony Vita, Wi-Fi Direct + Local, reduced latency, up-to-date, usercontrolled 4 Enables applications to focus on proximity instead of absolute location! Robin Kravets, University of Illinois
Won’t you be my neighbor? Detection Challenges Encounters are unplanned and unpredictable Nodes are energy-constrained Requires constant scanning Requires effective duty cycling Global Synchronization is difficult ? ? ? Requires asynchronous solutions Goal: Continuous Energy-efficient Asynchronous Neighbor Discovery 5 Robin Kravets, University of Illinois
Energy Efficiency: Duty-cycling Basic Discovery Idea 6 Time is slotted Nodes selectively remain awake for a full slot duration Nodes beacon at the beginning and end of an awake slot Discovery occurs when two active slots overlap Awake slots Robin Kravets, University of Illinois
Duty-cycled Neighbor Discovery Challenges: Dealing with unsynchronized slots Choosing active slots Dealing with asymmetric duty cycles Active Slot Selection Awake slots 7 Robin Kravets, University of Illinois
Slot Selection: Random Birthday protocol Randomly select a slot to wake up in with a given probability Advantage + Good average case performance Disadvantage - No bounds on worst-case discovery latency Fraction of Discoveries Cumulative Discovery Latency 8 Long tail Is a small delay bound really necessary? Average discovery → Useful contact time Good. Worst-case Avg. Case → Missed contacts Performance Discovery Latency Robin Kravets, University of Illinois
Slot Selection: Deterministic Disco (Sensys 2008) Each node selects two primes p 1 i and p 2 i Both nodes wake up every p 1 th and p 2 th slot (5 th and 7 th) Guarantees discovery in p 1 i x p 1 j slots U-Connect (IPSN 2010) 9 Both Disco and UConnect handle symmetric and asymmetric duty cycles Each node selects one prime pi Every node wakes up every pth slot and (p-1)/2 slots every p*p slots Overlap is guaranteed within pi x pj slots Robin Kravets, University of Illinois
Slot Selection: Deterministic Prime-based Strict worst-case bound Can we get the best of both worlds Good average discovery latency from random Poor average-case performance protocols Cumulative Discovery Latency Good delay bound from deterministic protocols Disadvantage 10 Advantage Fraction of Discoveries Disco U-Connect Birthday Discovery Latency Robin Kravets, University of Illinois
Searchlight Approach Have a deterministic discovery schedule that has a pseudo-random component Consider two nodes with the same (symmetric) duty cycles 3 slots Node A Node B A A B B Insight 11 Offset between slots with fixed period remains fixed Robin Kravets, University of Illinois
Searchlight Approach Have a deterministic discovery schedule that has a pseudo-random component Consider two nodes with the same (symmetric) duty cycles 4 slots Node A Node B Insight 12 A A B B 4 slots Offset between slots with fixed period remains fixed B will fall in the first t/2 slots of A’s cycle or A will fall in the first t/2 slots of B’s cycle Robin Kravets, University of Illinois
Searchlight Approach Have a deterministic discovery schedule that has a pseudo-random component Consider two nodes with the same (symmetric) duty cycles 4 slots Node A Node B Insight 13 A A B B 4 slots Offset between slots with fixed period remains fixed B will fall in the first t/2 slots of A’s cycle or A will fall in the first t/2 slots of B’s cycle Robin Kravets, University of Illinois
Systematic Probing Technique Select a fixed period t (does not need to be prime) Keep one slot fixed (anchor slot) t Node A A Node B B Add a second “probe” slot Objective is to meet the fixed/anchor slot of the other node Only need to search in the range 1 to t/2 No need to probe all t/2 slots all of the time 14 A Move around the probe slot Robin Kravets, University of Illinois
Sequential Probing 1 2 2 3 3 1 1 2 2 Discovery through anchor-probe overlap Two slots period t Guaranteed overlap in t*t/2 slots Anchor slot: Keep one slot fixed at slot 0 Probe slot: Move around the other slot sequentially Improved bound over existing protocols Based on the time needed to ensure a probe-anchor overlap But: Probe-probe overlap should also lead to discovery 15 Sequential scanning can result in probes “chasing” each other Robin Kravets, University of Illinois 3
Randomized Probing Break the pattern of chasing: 1 Move the probe slot randomly (A: 1 -3 -2; B: 3 -1 -2) 3 2 3 1 2 Discovery through probe-probe overlap Each node randomly chooses a schedule for its probe slot that repeats every (t*t/2) slots Schedules of two nodes appear random to each other Advantage 16 3 3 1 Pseudo-random instead of random 1 Retains the same worst-case bound Improves average case performance Robin Kravets, University of Illinois
Evaluation Comparison Protocols Birthday Disco U-Connect Fixed Energy Symmetric and asymmetric duty cycles Worst-case latency bound Cumulative discovery latency Methods Empirical and Simulation Implementation 17 All protocols operate at the same duty cycle Latency Sequential ( Searchlight-s) Random (Searchlight-r) Scenarios Metrics Searchlight Protocols Testbed of G 1 android and Nokia N 900 phones Robin Kravets, University of Illinois
Worst-case Latency Bound Metric: Energy Latency Product Protocol Disco U-Connect Searchlight 18 Duty Cycle Parameters Worstcase Latency Duty Cycle p 1, p 2 p t Robin Kravets, University of Illinois Worst-case bound for duty cycle 1/x Dutycycle for same bound
Worst-case Latency Bound Metric: Energy Latency Product Protocol Disco U-Connect Searchlight 19 Duty Cycle Parameters Worstcase Latency p 1, p 2 p 1 × p 2 p p 2 t t×(t/2) Duty Cycle Robin Kravets, University of Illinois Worst-case bound for duty cycle 1/x Dutycycle for same bound
Worst-case Latency Bound Metric: Energy Latency Product Protocol Disco U-Connect Searchlight 20 Duty Cycle Parameters Worstcase Latency Duty Cycle p 1, p 2 p 1 × p 2 4 x 2 p p 2 2. 25 x 2 t t×(t/2) 2 x 2 Robin Kravets, University of Illinois Worst-case bound for duty cycle 1/x Dutycycle for same bound
Worst-case Latency Bound Metric: Energy Latency Product Protocol Disco U-Connect Searchlight 21 Duty Cycle Parameters Worstcase Latency p 1, p 2 Duty Cycle Worst-case bound for duty cycle 1/x Dutycycle for same bound p 1 × p 2 4 x 2 2/x p p 2 2. 25 x 2 1. 5/x t t×(t/2) 2 x 2 1. 41/x Robin Kravets, University of Illinois
Symmetric Duty Cycles Fraction of Discoveries Cumulative Discovery Latency in Number of Slots 5% duty cycle 22 Robin Kravets, University of Illinois
Symmetric Duty Cycles Fraction of Discoveries Cumulative Discovery Latency in Number of Slots 5% duty cycle 23 Robin Kravets, University of Illinois
Symmetric Duty Cycles Fraction of Discoveries Cumulative Discovery Latency in Number of Slots 5% duty cycle 24 Robin Kravets, University of Illinois
Symmetric Duty Cycles Fraction of Discoveries Cumulative Discovery Latency in Number of Slots 5% duty cycle 25 Robin Kravets, University of Illinois
Symmetric Duty Cycles Fraction of Discoveries 820 960 Discovery Latency in Number of Slots Searchlight does not have the long tail of other deterministic protocols Searchlight-R performs almost as good as Birthday in the average case 26 Robin Kravets, University of Illinois
Handling Duty Cycle Asymmetry Why? Different energy requirements Different duty cycles (different values for t) Problem Anchor slots no longer have constant distance Node A (period=5) Node B (period=3) 29 Robin Kravets, University of Illinois
Handling Duty Cycle Asymmetry Solution Restrict choice of period to primes Overlap of anchor slots guaranteed through Chinese remainder theorem t needs to be prime Worst case latency is t 1 × t 2 Node A (period=5) Node B (period=3) 30 Robin Kravets, University of Illinois
Asymmetric (1% and 5%) Fraction of Discoveries Cumulative Discovery Latency 82% Discovery Latency in Number of Slots Searchlight-R 31 Worst-case latency is worse than both Disco and U-Connect Compensates for that by having best average case performance Robin Kravets, University of Illinois
Can we do better? Observation When slots are not fully aligned, slots of neighboring nodes overlap more than once within bound Anchor Slot Probe Slot 1 Probe Slot 2 Anchor Slot One overlap is sufficient for discovery! 32 Robin Kravets, University of Illinois
Striping across the rounds Insight Only need to probe alternate slots Anchor Slot Probe Slot 1 Probe Slot 2 Probe Slot 3 Probe Slot 4 Anchor Slot Reduces the number of active slots by almost ½! Problem 33 Slot alignment Robin Kravets, University of Illinois
Handling Slot Alignment 1 2 4 3 Probe Slot Anchor Slot 5 6 Probe Slot Anchor Slot δ Let the slots overflow a bit Extent of overlap ( ) depends on 34 Beacon transmission time Possible clock drift Robin Kravets, University of Illinois
Does it help? Protocol Disco U-Connect Searchlight Striped Searchlight 35 Duty Cycle Parameters Worstcase Latency p 1, p 2 p 1 × p 2 p p 2 t t×(t/2) t, δ t×(t/4) Duty Cycle Worst-case Duty-cycle bound for required for duty cycle same worst 1/x case bound δ = amount of “overflow” beyond regular slot boundary Robin Kravets, University of Illinois
Does it help? Protocol Disco U-Connect Searchlight Striped Searchlight 36 Duty Cycle Parameters Worstcase Latency Duty Cycle Worst-case Duty-cycle bound for required for duty cycle same worst 1/x case bound p 1, p 2 p 1 × p 2 4 x 2 p p 2 2. 25 x 2 t t×(t/2) 2 x 2 t, δ t×(t/4) (1+δ) 2 x 2 Robin Kravets, University of Illinois
Does it help? Protocol Disco U-Connect Searchlight Striped Searchlight 37 Duty Cycle Parameters Worstcase Latency Duty Cycle Worst-case Duty-cycle bound for required for duty cycle same worst 1/x case bound p 1, p 2 p 1 × p 2 4 x 2 2/x p p 2 2. 25 x 2 1. 5/x t t×(t/2) 2 x 2 1. 41/x t, δ t×(t/4) (1+δ) 2 x 2 (1+δ)/x Robin Kravets, University of Illinois
Striping and Asymmetry Problem Anchor slots no longer have constant distance Striping cannot be used Original approach Restrict choice of t to primes 38 Worst-case bound worse than other deterministic protocols Robin Kravets, University of Illinois
Maintaining Constant Offset New approach Restrict value of the bigger period to an integer multiple of the smaller period Node A (period=6) Node B (period=3) Other protocols also restrict the choice of values for their parameters 39 Only primes are allowed by Disco and U-Connect Robin Kravets, University of Illinois
Symmetric Duty Cycles Fraction of Discoveries Cumulative Discovery Latency in Number of Slots 5% duty cycle 40 Robin Kravets, University of Illinois Worst-case bound: 2000+ slots
Symmetric Duty Cycles Fraction of Discoveries Cumulative Discovery Latency in Number of Slots 5% duty cycle 41 Robin Kravets, University of Illinois Worst-case bound: 961 slots
Symmetric Duty Cycles Fraction of Discoveries Cumulative Discovery Latency Searchlight-S Discovery Latency in Number of Slots 5% duty cycle 42 Worst-case bound: 800 slots Robin Kravets, University of Illinois
Symmetric Duty Cycles Fraction of Discoveries Cumulative Discovery Latency Worst-case bound: 440 slots Searchlight-S Discovery Latency in Number of Slots 5% duty cycle Striped probing improves bound by almost 50% 43 Robin Kravets, University of Illinois
Fraction of Discoveries Asymmetric Duty Cycles Worst-case bound: 2266 slots Searchlight-S Discovery Latency in Number of Slots 1%-10% duty cycle 44 Robin Kravets, University of Illinois
Fraction of Discoveries Asymmetric Duty Cycles Worst-case bound: 1819 slots Searchlight-S Discovery Latency in Number of Slots 1%-10% duty cycle 45 Robin Kravets, University of Illinois
Fraction of Discoveries Asymmetric Duty Cycles Searchlight. S Discovery Latency in Number of Slots 1%-10% duty cycle Randomized probing does not have the same worst-case bound 46 Robin Kravets, University of Illinois
Restricted Randomized Probing Randomization across t. A/2 could delay discovery Node A (period=16) 3 2 1 Node B (period=8) Restrict randomization based on smallest t Impact 47 Same bound as sequential for asymmetric case No effect on symmetric case Robin Kravets, University of Illinois
What should I use? Mostly symmetric duty cycles Searchlight with restricted randomized striped probing For any two nodes with the same duty cycle For any two nodes with different duty cycles Best average and best worst-case bound Almost best average and best worst-case bound Very diverse duty cycles Searchlight with symmetric striped probing 49 Has slightly better average discovery latency Robin Kravets, University of Illinois
Searchlight: Won't You Be My Neighbor? http: //mobius. cs. uiuc. edu 50 Robin Kravets, University of Illinois
- Slides: 47