View
216
Download
0
Category
Preview:
Citation preview
Argos
Clayton W. ShepardHang Yu, Narendra Anand, Li
Erran Li, Thomas Marzetta, Richard Yang,
Lin Zhong
Practical Many-Antenna Base Stations
2
Motivation
• Spectrum is scarce
• Hardware is cheap
3
MU-MIMO Theory
• More antennas = more capacity
–Multiplexing
– Power
– Orthogonality
4
Data
2
Data 1
Data
6
Data 3
Data 4Data
5
5
Why not?
• Nothing scales with the number of antennas
– CSI Acquisition
– Computation
– Data Transportation
Background: Beamforming
6
=
Constructive Interference
=
Destructive Interference
?
7
Due to environment and terminal mobility estimation has to occur quickly and periodically
BS
The CSI is then calculated at the terminal and sent back to the BSA pilot is sent from each BS antenna
Background: Channel Estimation
+
+=
Align the phases at the receiver to ensure constructive interferenceFor uplink, send a pilot from theterminal then calculate CSI at BS
Path Effects (Walls)
Tx
Rx
Tx
Rx
Tx
Rx
Uplink?
Tx
Rx
Measured channels are not reciprocal due to differences in the Tx and Rx hardware!
8
Background: Multi-User Beamforming
Data 1
9
Background: Multi-User Beamforming
Data
2
10
Background: Null Steering
Data 1
Nu
ll
Null
NullNullN
ull
11
Background: Null Steering
Data
2
Null
NullNull
Null
Null
12
Background: Multi-User Beamforming
Data
2
Data 1
Data
6
Data 3
Data 4Data
5
13
Background: Scaling Up
Data 1
14
Background: Scaling Up
Data 1
15
Background: Scaling Up
Data 1
16
Data 3
Data
5
Background: Scaling Up
Data 1
Data
6
Data
2
Data 4
17
Background: Linear Precoding
• Calculate beamweights– Every antenna has a beamweight for
each terminal
• Multiply symbols by weight, then add together: sWs
18
Recap
1) Acquire CSI
2) Calculate Weights
3) Apply Linear Precoding
19
Scalability Challenges
1) Acquire CSI– M+K pilots, then M•K feedback
2) Calculate Weights– O(M•K2), non-parallelizable, centralized
data
3) Apply Linear Precoding– O(M•K), then O(M) data transport
20
Argos’ Solutions
1) Acquire CSI– New reciprocal calibration method
2) Calculate Weights– Novel distributed beamforming method
3) Apply Linear Precoding– Carefully designed scalable
architecture
O(M•K) → O(K)
O(M•K2) → O(K)*
O(M•K) → O(K) *
21
Solutions
• Reciprocal Calibration
• Distributed Beamforming
• Scalable Architecture
22
Channel Reciprocity
• Pilot transmission source?– Basestation– Terminal
• Base station pilot transmission – Requires feedback – M pilots (M ≥ K)
• Terminal pilot transmission– No feedback– K pilots
Channel Reciprocity
23
TxC
RxC
TxA
RxA
TxB
RxB
Channel Estimation
Transmission
Tx/Rx Chain differences
require calibration
Can we do this without terminal
involvement?
TxA+RxC-TxC-RxA
TxB+RxC-TxC-RxB
TxA+C+RxC-TxC-C-RxA
Key Idea
24
=
= Any constant phase shift results in same beampattern!
Channel Reciprocity
25
TxC
RxC
TxA
RxA
TxB
RxB
Channel Estimation
Transmission
TxA+RxC-TxC-RxA
TxB+RxC-TxC-RxB
TxA-RxA
TxB-RxB
• Find phase difference between A and B
• Tx from A:• Phase offset = TxA-RxA
• Tx from B: • (TxB-RxB )
Internal Reciprocal Calibration
26
TxA
RxA
TxB
RxB
TxA+C+RxB TxB+C+RxA TxA+RxB - TxB-
RxA
add phase difference + (TxA+RxB - TxB-RxA) = TxA-
RxA
27
Solutions
• Reciprocal Calibration
• Distributed Beamforming
• Scalable Architecture
28
Problems with Existing Methods
• Central data dependency
• Transport latency causes capacity loss
• Can not scale– Becomes exorbitantly expensive then
infeasible
29
Conjugate Beamforming
• Requires global power scaling by constant:
• Where, e.g.:
• This creates a central data dependency
30
BS
Good Channel
Conjugate Beamforming Power
Okay Channel
Bad Channel
31
BS
High Power
Conjugate Beamforming Power
Normal Power
Low Power
32
Distributed Conjugate Beamforming
• Scale power at each antenna:
• Maximizes utilization of every radio–More appropriate for real-world deployments
• Quickly approaches optimal as K increases– Channels are independent and uncorrelated
33
BS
High Power
Distributed Conjugate Beamforming
High Power
High Power
34
Solutions
• Reciprocal Calibration
• Distributed Beamforming
• Scalable Architecture
?
• Scalable– Support
thousands of BS antennas
• Cost-effective– Cost scales
linearly with # of antennas
35
Architectural Design Goals…
Data
36
Linear Precoding…
M
K
K
K
K
…
…
…
…
37
Scalable Linear Precoding…
M
K
K
K
K
…
…
…
…
38
Scalable Linear Precoding…
M
K
K
K
K
Common Databus!
…
…
…
…
39
MUBF Linear Precoding: Uplink
…
M
K
…
K
K
K
…
…
…
40
Scalable Linear Precoding
…
…
…
…
…
M
K
Constant Bandwidth!
K
K
K
41
Partition Ramifications
• CSI and weights are computed and applied locally at each BS radio– No overhead for additional BS radios
• No central data dependency– No latency from data transport– Constant data rate common bus (no
switching!)
• Unlimited scalability!
42
How do we design it?
• Daisy-chain (series)– Unreliable– Large end to end latency
• Token-ring | Interconnected– Not amenable to linear precoding– Variable Latency– Routing overhead
…
…
…
• Flat structure– Unscalable– Expensive, with large fixed cost
43
Solution: Argos Architecture
Central Controller
Argos Hub
Argos Hub
Argos Hub
Module
Module
Module
Module
Module
…
…
…
Data Backhaul
ModuleRadio Radio Radio…
44
Argos Implementation
Central Controller
Argos Hub
Module
Module
Module
…
Central Controller
(PC with MATLAB)
Sync Pulse
Ethernet
Clock Distribution
Argos Hub Argos
Interconnect
WARP Module
FPGA
Power PC
FPGA Fabric
Hardware Model
Peripherals and Other
I/O
Clock Board
Daughter
Cards
Radio 4
Radio 3
Radio 2
Radio 1
WARP Module
FPGA
Power PC
FPGA Fabric
Hardware Model
Peripherals and Other
I/O
Clock Board
Daughter
Cards
Radio 4
Radio 3
Radio 2
Radio 1
16
………
Eth
ern
et
WARP Module
FPGA
Power PC
FPGA Fabric
Hardware Model
Peripherals and Other
I/O
Clock Board
Daughter
Cards
Radio 4
Radio 3
Radio 2
Radio 1
ArgosInterconne
ct
45
46
WARP Module
s
Central Controll
er
Argos Hub
Clock Distributio
n
Ethernet
Switch
SyncDistributio
n
Argos Interconnects
47
System Performance
48
Linear Gains as # BS Ant. Increases
Capacity vs. M, with K = 15
49
Linear Gains as # of Users Increases
Capacity vs. K, with M = 64
50
Scaling # of Users with 16 BS Ant.
Capacity vs. K, with M = 16
51
Zero-forcing is not always better!
Capacity vs. K, with M = 16 | Low Power
52
Calibration is stable for hours!
53
Conclusion
• First many-antenna beamforming platform– Real-world demonstration of manyfold
capacity increase
• Devised novel techniques and architecture– Unlimited Scalability
http://recg.rice.edu
http://argos.rice.edu
Recommended