Improving Cellular Capacity with White Space Offloading

Transcription

1 Improving Cellular Capacity with White Space Offloading Suzan Bayhan*, Liang Zheng*, Jiasi Chen, Mario Di Francesco, Jussi Kangasharju, and Mung Chiang * equal contribution WiOPT, Paris, France, May 15-19,

2 Motivation Increasing wireless traffic Limited cellular capacity Cisco Forecasts 49 Exabytes per Month of Mobile Data Traffic by

3 Motivation Offloading cellular traffic to white spaces 3

4 Motivation Offloading cellular traffic to white spaces Frequencies that are spatio-temporally unused despite being allocated to primary users (PUs) and that can be used opportunistically by secondary users. 4

5 TV white spaces (TVWS) Frequencies in UHF and VHF (~ MHz) In 2010, FCC made this spectrum open for unlicensed public access via a spectrum database (WSDB) Wide coverage range in the order of kms Better penetration through walls and vegetation 5

6 Arising questions How much can we offload to white spaces (in the downlink)? How to choose users for white space offloading? Which white space channels they should connect to? Overall network throughput Individual user s throughput Switching cost 6

7 Our approach Model offloading problem as an optimization of resource allocation Design solutions for white-space offloading driven by the insights from formulated optimization problem Evaluation using simulations and real data from Google Spectrum Database 7

8 System model Inputs: User locations Rayleigh channel states WSDB availability data Network controller offloading decision every timeslot (L ws channels) WSDB spectrum info Dual-mode BSs Users in cellular User in white spaces Dual-mode users 8

9 System model Inputs: User locations Rayleigh channel states WSDB availability data Network controller offloading decision every timeslot (L ws channels) WSDB spectrum info Dual-mode BSs Cellular resources: owned by the cellular operator and cell frequencies orthogonally assigned Dual-mode users White spaces: Accessed after query to a WSDB, subject to multi-user Users in cellular interference User in white spaces 9

10 Cellular throughput Power of BSk at user i: pi k Number of cellular users at BS k: Nc k Bandwidth sharing 10

11 White space throughput Bandwidth and power according to the WSDB response Defined by the regulations 11

12 Our approach Model offloading problem as an optimization of resource allocation Design solutions for white-space offloading driven by the insights from formulated optimization problem Evaluation using simulations and real data from Google Spectrum Database 12

13 Joint optimization of power and offloading 13

14 Joint optimization of power and offloading Offloading decision: set of users to offload to white spaces 14

15 Joint optimization of power and offloading Power allocation decision: transmission power of BSk for each of its users 15

16 Joint optimization of power and offloading Cellular aggregate throughput 16

17 Joint optimization of power and offloading Cellular aggregate throughput white space throughput 17

18 Joint optimization of power and offloading Cellular aggregate throughput white space throughput Cellular users: total power constraints in a multi-user downlink system White space users: fixed power limit set by the regulators and recorded in the WSDB response message. 18

19 Optimal cellular tx power Suppose a given offloading decision From KKT conditions, we obtain the optimal tx powers for the cellular users 19

20 Optimal cellular tx power Suppose a given offloading decision From KKT conditions, we obtain the optimal tx powers for the cellular users Power shared almost equally 20

21 Optimal cellular tx power Suppose a given offloading decision From KKT conditions, we obtain the optimal tx powers for the cellular users Power shared almost equally Normalized in a way such that less power is allocated to users with a poorer channel condition 21

22 Optimal offloading decision Given the optimal power allocation, we derive the optimal offloading decision Based on complementary slackness, we find necessary conditions for optimality 22

23 Optimal offloading decision Given the optimal power allocation, we derive the optimal offloading decision Based on complementary slackness, we find necessary conditions for optimality Gain in the user s throughput 23

24 Optimal offloading decision Given the optimal power allocation, we derive the optimal offloading decision Based on complementary slackness, we find necessary conditions for optimality Gain in the user s throughput RHS: Average gain in throughput for all users in the network. 24

25 Low-complexity scheme: Positive Gain (PG) PG considers how the assignment of a given user n to a specific channel affects the throughput of the users already allocated to that channel. Users are switched to white space if: Increase in overall network throughput Increase in this user s throughput 25

26 Switching overhead To avoid frequent interface switching, we define a cost function reflecting the switching overhead/latency (and associated throughput loss) Switching among white space channels: cost linear increasing function of frequency separation [Gozupek 2013] Interface switching: constant cost higher than intra-white space switching D. Gozupek, S. Buhari, and F. Alagoz, A spectrum switching delay-aware scheduling algorithm for centralized CRNs, IEEE Trans. Mobile Comput., vol. 12, no. 7, pp ,

27 Switching overhead To avoid frequent interface switching, we define a cost function reflecting the switching overhead/latency (and associated throughput loss) Switching among white space channels: cost linear increasing function of frequency separation [Gozupek 2013] Interface switching: constant cost higher than intra-white space switching While calculating expected throughput, Update expected throughput as: (1-switching overhead)*throughput D. Gozupek, S. Buhari, and F. Alagoz, A spectrum switching delay-aware scheduling algorithm for centralized CRNs, IEEE Trans. Mobile Comput., vol. 12, no. 7, pp ,

28 Switching-cost aware offloading (SCA) Expected switching latency calculated using the expected availability probability of white space channels and staying time at an interface If latency below some threshold, offload the user to that white space channel 28

29 Our approach Model resource allocation in the downlink Design solutions for white-space offloading driven by the insights from formulated optimization problem Evaluation using simulations and real data from Google Spectrum Database 29

30 Performance evaluation Impact of User Density Impact of Cellular Channel Quality Impact of Switching cost Google spectrum database 30

31 Performance evaluation Cellular (CELL): no white space offloading Random Interface (RI): select one interface randomly Greedy (GR): Each user is assigned to the interface with the highest throughput (impact of new user on existing ones NOT considered) Switching Cost-Aware (SCA) Positive Gain (PG) 31

32 Scenarios: two cities 4km x 4 km area in 2 US cities: white space availability depending on the population-density: 1.New York City (NYC) 2.Boulder in Colorado (BC) Base station deployment from OpenSignal crowdsourcing errors, merge into 100 BS for NYC, 35 BS in BC 32

33 Scenarios: spectrum availability - Google WSDB: no temporal variability for the same region Aleksandr Zavadovski, Suzan Bayhan, and Jussi Kangasharju, Data-Driven Analysis of DatabaseAssisted Spectrum Access for Mobile Users, INFOCOM Workshop on SCAN 2017 NYC: no channel available in downtown, 1 ch in wider region assumed availability ~ U(0.95, 1) Boulder in Colorado (BC): 4 chs. available U(0.7,1) for BC 33

34 Cellular vs. White space setting 40 W for cellular, 4 W for white spaces 10 MHz for cellular, 6 MHz for white spaces Same Rayleigh channel gain with average value 1 Channel switching latency s/6 MHz Users are mobile with U(0,10) m/s 100 time slots, average of 10 runs 34

35 Impact of user density Boulder NYC PG(30%-62%), GR, SCA provide a significant additional capacity over CELL PG attains the same capacity under increasing density SCA trade-off capacity for stability GR degrades with more users 35

36 Impact of user density Boulder NYC For NYC, gain is less pronounced. Yet, around Mbps/cell corresponding to an 8 17% improvement over CELL. 36

37 Boulder How much offloaded to white spaces? NYC Offloaded to white spaces Up to 37% traffic offloaded to white spaces in BC ~13% in NYC 37

38 Impact of the cellular channel quality Assume indoor users white space better penetration through walls Cellular channel gain for indoors (0.01) White space channel gains in/out-doors (0.1) Cellular channel gain for outdoors (0.1) Lower cellular channel gain for indoors 38

39 Impact of the cellular channel quality Increase in capacity with white space offloading Traffic offloaded to white spaces 39

40 Impact of switching cost PG and GR sustain the throughput by considering switching cost SCA approaches to CELL as under high cost, it does not use white space channels. 40

41 Summary How much capacity can we harness? - highly depends on the considered scenarios! - white space availability - population density - cellular capacity increase by up to 16% in NYC, 62% in Boulder White spaces significantly benefit indoor users As future work: - adopt existing white space power allocation scheme - power efficiency as another decision parameter 41

42 Thank you! 42

43 Proof of Proposition 1 White space power already set Offloading decision is done How to set the cellular powers? 43

44 Interface switching: stability analysis To avoid frequent interface switching, we define a cost function reflecting the switching overhead/latency (and associated throughput loss) Interface switching: constant cost higher than intra-white space switching White space switching: cost linear increasing function of frequency separation 44

45 Metrics Total capacity per cell as the aggregate throughput in a cell achieved by assigning all users to different resources. Per user throughput as the average throughput per user. Fraction of traffic through each interface as the ratio of the traffic served through a particular interface (i.e., either cellular or white space) to the total capacity of a cell. Probability of frequency (interface) switching as the probability that a user will be assigned to a different frequency (interface) in the next time slot. White space re-use distance as the average minimum distance between users on the same white space channel. 45