Practical Conflict Graphs for Dynamic Spectrum Distribution Xia

Practical Conflict Graphs for Dynamic Spectrum Distribution Xia Zhou, Zengbin Zhang, Gang Wang, Xiaoxiao Yu*, Ben Y. Zhao and Haitao Zheng Department of Computer Science, UC Santa Barbara *Tsinghua University, China

Inefficient Spectrum Distribution • Explosive wireless traffic growth • The well-know problem: artificial spectrum shortage – Spectrum is assigned statically – Hard to get new spectrum – Current spectrum utilization is low Need efficient spectrum distribution 2

Dynamic Spectrum Distribution Spectrum • Key requirements – Reuse spectrum in space whenever possible ? A ? B – Exclusive spectrum access for allocated users ? C Must characterize interference conditions among users 3

Conflict Graphs • Binary representation of pairwise interference conditions Coverage area: all receiver locations A A C B B C 4

Benefits of Conflict Graphs • Simple abstraction – Reduce spectrum allocation to graph coloring problems • Leverage numerous graph algorithms – Many efficient allocation algorithms • Widely used 5

Key Issues on Conflict Graphs • Hard to get it accurate #1 – Wireless propagation is complex – Exhaustive measurements are not scalable – Solutions w/o measurements give errors, poor performance • Fail to capture accumulative interference #2 – A fundamental graph limitation – Interference cumulate from multiple transmissions C A B Are conflict graphs useful in practice? 6

Overview • Goal: understand practical usability of conflict graphs • Contributions – A practical method of building conflict graphs – Measurement validation of graph accuracy – Graph augmentation to address accumulative interference 7

Outline • Introduction • Measurement-Calibrated Conflict Graphs • Validation Results • Graph Augmentation 8

Building Practical Conflict Graphs Exhaustive measurements Accurac y Non-measurement methods Our Goal Measurement overhead • Our approach: measurement-calibrated conflict graphs 9

Measurement-Calibrated Conflict Graphs Monitor Sampled Signal Measurements Calibrated Propagation Model ? Exhaustive Signal Measurements Measured Conflict Graph Predicted Signal Maps Estimated Conflict Graph 10

Evaluating Conflict Graphs • Compare estimated and measured conflict graphs Monitor Exhaustive Signal Measurements Signal Prediction Accuracy Sampled Signal Measurements Calibrated Propagation Model Predicted Signal Maps Measured Conflict Graph Similarity Spectrum Allocation Results Spectrum Allocation Benchmark s Estimated Conflict Graph Spectrum Allocation Results 11

Measurement Datasets • Exhaustive signal measurements at outdoor Wi. Fi networks Area Avg # of APs #of (km 2 heard per APs ) location # of measured locations Dataset Location Metro. Fi Portland, OR 7 70 2. 3 30, 991 TFA Network Houston, TX 3 22 2. 7 27, 855 7 78 6. 2 11, 447 Google. Wi. F Mountain View, CA i • Our own dataset collected at Google. Wifi – Capture weak signals using radio with higher sensitivity 12

Outline • Introduction • Measurement-Calibrated Conflict Graphs • Validation Results • Graph Augmentation 13

Evaluating Conflict Graphs Exhaustive Signal Measurements Signal Predictio n Accuracy Predicted Signal Maps Measured Conflict Graph Similarity Estimated Conflict Graph Spectrum Allocation Results Spectrum Allocation Benchmarks Spectrum Allocation Results 14

Signal Prediction Results • Predict signal values using a sample of measurements – Models: Uniform, Two-Ray, Terrain, and Street – Street model achieves the best accuracy # of occurrences • Location-dependent pattern in prediction errors 4000 Underprediction Overprediction Underpredict RSS values at closer locations 3000 2000 Overpredict RSS values at farther locations 1000 0 0 0, 1 0, 2 0, 3 0, 4 Distance to AP (km) 0, 5 15

Evaluating Conflict Graphs Exhaustive Signal Measurements Signal Predictio n Accuracy Predicted Signal Maps Measured Conflict Graph Similarity Estimated Conflict Graph Spectrum Allocation Results Spectrum Allocation Benchmarks Spectrum Allocation Results 16

Conflict Graph Accuracy • Extra edge: in estimated graph but not measured graph • Missing edge: in measured graph but not estimated graph Extra edges dominate! Correct edge Extra edge Missing edge 17

Why Do Extra Edges Dominate? • Signal prediction errors are location-dependent – An edge exists if Signal-to-Interference-and-Noise Ratios (SINRs) < a threshold SINR = Signal Interferenc + Noise e Under-estimate receivers’ SINR values more conflict edges 18

Evaluating Conflict Graphs Exhaustive Signal Measurements Signal Predictio n Accuracy Predicted Signal Maps Measured Conflict Graph Similarity Estimated Conflict Graph Spectrum Allocation Results Spectrum Allocation Benchmarks Utilization Reliability Spectrum Allocation Results 19

Spectrum Allocation Benchmarks • Estimated graphs are conservative • Estimated graphs has lower spectrum utilization – Utilization: spectrum reuse • Estimated graph has higher reliability – Reliability: % of users receive reliable spectrum use – Still, users suffer accumulative interference Need to address accumulative interference! 20

Graph Augmentation • Key idea: add edges selectively to improve reliability • Our solution: greedy augmentation – Integrate spectrum allocation to identify edges to add – More details in the paper • Result: 96%+ users receive reliable spectrum use 21

Our Conclusion: Conflict Graphs Work! 22

BACKUP 23

Collecting Google. Wifi Dataset • 3 -day wardriving • 3 co-located laptops, each monitoring one channel • Locations have 5 m separation on average 24

Impact of Sampling Rate • 34 monitors per km 2 achieve the best tradeoff for the urban street environment • Determine sampling rate – Depends on AP density, propagation environment, and monitor’s sensitivity 25

Signal Prediction Errors • Errors are noticeable, Gaussian distribution – Align with prior studies 26

Building Conflict Graphs • Coverage-based conflict graph – Node: a spectrum user with its coverage region – Edge: e. AB exists if when A and B use the same channel, A or B fails to maintain γ of its receptions successful A Signal S ce I en Interfer B Reception succeeds if SINR is above a threshold 27

Spectrum Allocation Benchmarks • Allocation algorithm – Multi-channel allocation: maximize proportional fairness • Metric #1: spectrum efficiency – Average fraction of spectrum received per user Extraneous edges lead to moderate efficiency loss (< 30%) 28

Spectrum Allocation Benchmarks • Metric #2: spectrum reliability – Fraction of users with exclusive spectrum usage – Consider interference from all the others on the same channel Extraneous edges reduce the impact of accumulative interference Need to address accumulative interference! 29

Graph Augmentation Results • Augmentation improves graph accuracy – Some edges added in measured graph are already in estimated graph 30

Efficacy of Graph Augmentation • Address accumulative interference – Eliminate reliability violations for measured graphs – 96+% reliability for estimated graphs – Add minimal edges, leading to efficiency loss < 15% for estimated graph 31