Argos: Practical Base Stations for Large-scale Beamforming. Clayton W. Shepard

Transcription

1 Argos: Practical Base Stations for Large-scale Beamforming Clayton W. Shepard

2 Collaborators Hang Yu Narendra Anand Erran Li Thomas Marzetta Richard Yang Lin Zhong 2

3 = Background Beamforming Power Gain Adjust phase ( beamweights ) Leverages Interference = Open-loop Pre-compute weights to specify direction Closed-loop (adaptive) Use channel state information (CSI) to target receivers 3

4 Background Single-user beamforming (SUBF) W SUBF c H * Multi-user beamforming (MUBF) W MUBF c H * ( H T H * ) 1 4

5 Background: Channel Estimation For Due Path uplink, to Effects environment send (Walls) a pilot and from terminal the The Align CSI the is phases calculated the receiver at the to A terminal mobility pilot is sent estimation then Uplink? from each has CSI to BS occur antenna at BS terminal ensure (Channels quickly constructive and are periodically sent notback reciprocal) interference to the BS Tx Rx Tx BS Rx Tx + + = Tx Rx Rx 5

6 MUBF linear pre-coding: downlink K K K K M 6

7 MUBF linear pre-coding: uplink K K K K M 7

8 Our vision 8

9 Prior Work Large-scale beamforming theory T.L. Marzetta. Noncooperative Cellular Wireless with Unlimited Numbers of Base Station Antennas. IEEE Transactions on Wireless Communications, Nov Fredrik Rusek and Daniel Persson and Buon Kiong Lau and Erik G. Larsson and Thomas L. Marzetta and Ove Edfors and Fredrik Tufvesson Scaling up MIMO: Opportunities and Challenges with Very Large Arrays. arxiv, Jan Real-world beamforming E. Aryafar, N. Anand, T. Salonidis, and E. Knightly. Design and Experimental Evaluation of Multi-userBeamforming in Wireless LANs. In Proceedings of MobiCom, 2010 Reciprocal calibration F. Kaltenberger, H. Jiang, M. Guillaud, R. Knopp. Relative channel reciprocity calibration in MIMO/TDD systems. Future Network and Mobile Summit, June

10 First large-scale beamforming base station 10

11 11

12 12

13 Overview of contributions Scalable architecture Internal reciprocity calibration Novel fully distributed beamforming method 13

14 Can beamforming scale with the number of base station antennas? 14

15 Not with current techniques! CSI acquisition Typically requires # of base station (BS) antennas (M) + # of terminals (K) pilots Weight calculation All existing methods have centralized data dependency Requires M*K channel estimates and produces M*K weight values Linear pre-coding Produces M data streams 15

16 With careful design and new techniques it can! CSI Acquisition Leverage TDD reciprocity to limit pilots to K Requires calibration Weight Calculation Novel decentralized weight calculation Linear Pre-coding Apply weights at radio For uplink combine streams any time they meet 16

17 Scalable linear pre-coding Common Databus! K K K K M 17

18 MUBF linear pre-coding: uplink K K K K M 18

19 Scalable linear pre-coding Constant Bandwidth! K K K K M 19

20 Ramifications CSI and weights are computed and applied (linear pre-coding) locally at each BS radio No overhead for additional BS radios No central data dependency No latency from data transport No stringent latency requirements Constant data rate common bus (no switching!) Unlimited scalability! 20

21 Design goals Scalable Support thousands of BS antennas??? Cost-effective Cost scales linearly with # of antennas Reliable 21

22 How do we design it? Daisy-chain (series) Unreliable Large end to end latency Flat structure Un-scalable Expensive, with large fixed cost Token-ring / Interconnected Not amenable to linear pre-coding Variable Latency Routing overhead 22

23 Solution: Argos Modular Daisy-chainable 1 or more radios Central Controller Argos Hub Argos Hub Data Backhaul Argos Hub Hierarchal Increases Reliability Constrains Latency Cost-effective Module Module Module Module Module Module Radio Radio Radio 23

24 Scalability of Argos Scalable in 4 directions: # of Radios per Module # of Modules per Chain # of ports per Hub # of Hubs (and levels) Reliable Branches can fail without affecting other branches Central hubs can be easily made redundant Accommodates linear pre-coding Add samples together at every junction 24

25 Ethernet Implementation Central Controller (Host PC w/matlab) Argos Hub Data Switch (Ethernet) Sync Pulse (WARP Board) Clock Distribution (AD9523) FPGA FPGA FPGA FPGA (controlled by XPS) (controlled by XPS) (controlled Power by XPS) PC (controlled Power (C by code XPS) PC Target) Power (C code PC Target) Power (C code FPGA PC Target) Fabric (C code FPGA Target) Fabric Argos Argos Argos Peripherals FPGA and Fabric Hardware Model Argos Interconnect Interconnect Interconnect Peripherals FPGA Other and I/O Fabric Hardware (SimuLink) Model Interconnect Peripherals Other and I/O Hardware (SimuLink) Model Peripherals Other and I/O Hardware (SimuLink) Model Other I/O (SimuLink) Clock Board Clock Board Clock Board Clock Board WARP Module WARP Module WARP Module Daughter WARP Module Daughter Cards Daughter Cards Daughter Cards Radio 1 Cards Radio 1 Radio 1 Radio 1Radio 2 Radio 2 Radio 2 Radio 2Radio 3 Radio 3 Radio 3 Radio 3Radio 4 Radio 4 Radio 4 Radio

26 Central Controller WARP Modules Argos Interconnects Sync Distribution Argos Hub Clock Distribution Ethernet Switch 26

27 Overview of contributions Scalable architecture Internal reciprocity calibration Novel fully distributed beamforming method 27

28 Channel reciprocity h i j tx i crx j tx i hi j rx j c h ji tx j crx i rx i hj i tx j Transciever i Wireless Channel Transciever j 28

29 Calibration coefficients Given the complete channel: We define a calibration coefficient as: Thus: i j j i j i h h A j i j i rx c tx h i j j i j i h A h i j j i A A A 1 1 and 29 i j i j j i A rx tx rx tx 1 i j j i rx c tx rx c tx

30 Applying to large-scale BS Find A between each BS antenna and a reference antenna (1) A 1m Every BS radio listens to terminal pilot ht m Find A between reference and terminal A 1t We can derive A mt A A 1t 1m Now every h can be found via h mt A mt h tm 30

31 Key observation But this requires K+1 pilots Even worse, it requires feedback A constant phase shift across the entire array does not alter the beampattern! h mt A mt h tm A A 1t 1m h tm 1 A 1m h tm Assuming A1 t 1 results in a constant phase offset, and thus does not affect radiation pattern 31

32 Internal calibration We find all A 1m offline They are static, and can be found quickly Send K orthogonal pilots to find all Used for uplink beamforming directly h m t k Use h h tm mt A1 m for downlink beamforming 32

33 Overview of contributions Scalable architecture Internal reciprocity calibration Novel fully distributed beamforming method 33

34 Problem with existing methods Central data dependency Transport latency causes capacity loss Can not scale Becomes exorbitantly expensive then infeasible 34

35 Conjugate beamforming Requires global power scaling by constant: Where, e.g.: This creates a central data dependency 35

36 Local conjugate beamforming Scale power locally: Maximizes utilization of every radio More appropriate for real-world deployments Quickly approaches optimal as K increases Channels are independent and uncorrelated 36

37 Capacity Gain Results Huge Capacity Gains Capacity Gain for M = 64 Local Conj. Global Conj. 8 Zeroforcing Performance linear with M and K 0 Same Power 1/64th Power Channel Calibration Stable Local conjugate indistinguishable from global Approaches optimality quickly with K 37

38 Results: scaling M Capacity vs. M, with K = 15 38

39 Results: scaling K Capacity vs. K, with M = 64 39

40 Results: scaling K Capacity vs. K, with M = 16 40

41 Results: low power Capacity vs. K, with M = 16 41

42 Results: calibration stability 42

43 Results: local conjugate 43

44 Future directions Find optimal tradeoff between zeroforcing and conjugate Demonstrate network optimality Lower power reduces other-cell interference Leverage cooperative beamforming Investigate promising match with full duplex Leverage huge EIRP gains 44

45 Conclusion First large-scale beamforming platform Real-world demonstration of manyfold capacity increase Devised novel architecture and techniques Unlimited Scalability 45

46 Acknowledgements Theoretical Discussion and Background Ashutosh Sabharwal WARP Support Patrick Murphy, Gaurav Patel, Chris Hunter, Sidharth Gupta Platform Construction Nathan Zuege, Chris Harris, Azalia Mirhoseini, Danny Eaton, Paul Williams 46