dsl management
Embed Size (px)
TRANSCRIPT
-
8/6/2019 Dsl Management
1/42
DSL Spectrum Management
Dr. Jianwei Huang
Department of Electrical EngineeringPrinceton University
Guest Lecture of ELE539A
March 2007
Jianwei Huang (Princeton) DSL Spectrum Management March 2007 1 / 26
-
8/6/2019 Dsl Management
2/42
Acknowledgements
Collaborations: Raphael Cendrillon, Mung Chiang, Marc MoonenSponsorships: Alcatel, NSF
Jianwei Huang (Princeton) DSL Spectrum Management March 2007 2 / 26
-
8/6/2019 Dsl Management
3/42
Digitial Subscriber Line (DSL) Networks
Wireline communications networks based telephone copper lines
Cost-effective broadband access networkMore than 160 million users world-wide
crosstalk
TX
TX RX
RXCO
RT
(Remote Terminal)
(Central Office) Customer
Customer
Jianwei Huang (Princeton) DSL Spectrum Management March 2007 3 / 26
-
8/6/2019 Dsl Management
4/42
Digitial Subscriber Line (DSL) Networks
Wireline communications networks based telephone copper lines
Cost-effective broadband access networkMore than 160 million users world-wideSpeed is the bottleneck
crosstalk
TX
TX RX
RXCO
RT
(Remote Terminal)
(Central Office) Customer
Customer
Jianwei Huang (Princeton) DSL Spectrum Management March 2007 3 / 26
-
8/6/2019 Dsl Management
5/42
How DSL Works?
Copper line can support signal transmissions over a large bandwidthVoice transmission: up to 3.4 KHzDSL transmissions: up to 30 MHz
Multi-carrier transmissions: Discrete Multitone Modulation
Frequency (KHz)0 3.4
Voice DSL
Jianwei Huang (Princeton) DSL Spectrum Management March 2007 4 / 26
-
8/6/2019 Dsl Management
6/42
Network and Channel Model
crosstalk
TX
TX RX
RXCO
RT
(Remote Terminal)
(Central Office) Customer
Customer
Jianwei Huang (Princeton) DSL Spectrum Management March 2007 5 / 26
Mathematical model : multi-user multi-carrier interference channelEach telephone line is a user (transmitter-receiver pair)
Generate mutual crosstalks over multiple frequency tones
-
8/6/2019 Dsl Management
7/42
Network and Channel Model
crosstalk
TX
TX RX
RXCO
RT
(Remote Terminal)
(Central Office) Customer
Customer
Jianwei Huang (Princeton) DSL Spectrum Management March 2007 5 / 26
Physical model : mixed CO/RT caseChannel attenuates with distance
Central Office (CO) connect customers who are reasonably closeRemote Terminal (RT) connect customers who are farther away
-
8/6/2019 Dsl Management
8/42
Network and Channel Model
crosstalk
TX
TX RX
RXCO
RT
(Remote Terminal)
(Central Office) Customer
Customer
Jianwei Huang (Princeton) DSL Spectrum Management March 2007 5 / 26
Frequency-Dependent ChannelDirect channel gain decreases with frequency
Crosstalk channel gain increases with frequency
-
8/6/2019 Dsl Management
9/42
Network and Channel Model
crosstalk
TX
TX RX
RXCO
RT
(Remote Terminal)
(Central Office) Customer
Customer
Jianwei Huang (Princeton) DSL Spectrum Management March 2007 5 / 26
Frequency-Dependent ChannelDirect channel gain decreases with frequency
Crosstalk channel gain increases with frequencyLead to near-far problem
RT generates strong crosstalk to CO line, especially inhigh tonesCO generates little crosstalk to RT in all tones
-
8/6/2019 Dsl Management
10/42
Crosstalk System Model
N users (lines) and K tones (frequency bands)User ns achievable rate on tone k is
b k n = log 1 + SINRk n
whereSINR k n =
p k nm = n k n , m p k m + k n
Total data rate of user n
R n =k
b k n
Jianwei Huang (Princeton) DSL Spectrum Management March 2007 6 / 26
-
8/6/2019 Dsl Management
11/42
Network Objective: Maximize Rate Region
Jianwei Huang (Princeton) DSL Spectrum Management March 2007 7 / 26
Rate Region : set of all achievable rate vectors
1
R
Rate Region
2
R
-
8/6/2019 Dsl Management
12/42
Network Objective: Maximize Rate Region
Jianwei Huang (Princeton) DSL Spectrum Management March 2007 7 / 26
Problem A: (Find One Point On the Rate Region Boundary)maximize
{pnP n }n nw n R n
User n chooses a power vector pn P n = k p k n P maxn , p k n 0 .
Rate Region : set of all achievable rate vectors
1
R
Rate Region
2
R
-
8/6/2019 Dsl Management
13/42
Network Objective: Maximize Rate Region
Jianwei Huang (Princeton) DSL Spectrum Management March 2007 7 / 26
Problem A: (Find One Point On the Rate Region Boundary)maximize
{pnP n }n nw n R n
User n chooses a power vector pn P n = k p k n P maxn , p k n 0 .Changing different weights trace the entire rate region boundary
Rate Region : set of all achievable rate vectors
1
R
Rate Region
2
R
-
8/6/2019 Dsl Management
14/42
Network Objective: Maximize Rate Region
Jianwei Huang (Princeton) DSL Spectrum Management March 2007 7 / 26
Problem A: (Find One Point On the Rate Region Boundary)maximize
{pnP n }n nw n R n
User n chooses a power vector pn P n = k p k n P maxn , p k n 0 .Changing different weights trace the entire rate region boundaryA suboptimal algorithm leads to a reduced rate region
Rate Region : set of all achievable rate vectors
R
Rate Region
2
R1
-
8/6/2019 Dsl Management
15/42
Difficulties of Solving Problem A
Non-convexity: total weighted rate not concave in power.
Physically distributed: local channel information
Performance coupling: across users (interferences) and tones (powerconstraint)
Jianwei Huang (Princeton) DSL Spectrum Management March 2007 8 / 26
-
8/6/2019 Dsl Management
16/42
Dynamic Spectrum Management (DSM)State-of-art DSM algorithms:
IW: Iterative Water-lling [Yu, Ginis, Cioffi02]
IW
2
R 1
R
Algorithm Operation Complexity PerformanceIW Autonomous O (KN ) Suboptimal
Jianwei Huang (Princeton) DSL Spectrum Management March 2007 9 / 26
-
8/6/2019 Dsl Management
17/42
Dynamic Spectrum Management (DSM)State-of-art DSM algorithms:
IW: Iterative Water-lling [Yu, Ginis, Cioffi02]OSB: Optimal Spectrum Balancing [Cendrillon et al.04]ISB: Iterative Spectrum Balancing [Liu, Yu05] [Cendrillon, Moonen05]
OSB/ISB
IW
2
R 1
R
Algorithm Operation Complexity PerformanceIW Autonomous O (KN ) SuboptimalOSB Centralized O Ke N OptimalISB Centralized O KN 2 Near Optimal
Jianwei Huang (Princeton) DSL Spectrum Management March 2007 9 / 26
-
8/6/2019 Dsl Management
18/42
Dynamic Spectrum Management (DSM)State-of-art DSM algorithms:
IW: Iterative Water-lling [Yu, Ginis, Cioffi02]OSB: Optimal Spectrum Balancing [Cendrillon et al.04]ISB: Iterative Spectrum Balancing [Liu, Yu05] [Cendrillon, Moonen05]ASB: Autonomous Spectrum Balancing [Huang et al.06]
/ASBOSB/ISB
IW
R
1 R
2
Algorithm Operation Complexity PerformanceIW Autonomous O (KN ) SuboptimalOSB Centralized O Ke N OptimalISB Centralized O KN 2 Near OptimalASB Autonomous O (KN ) Near Optimal
Jianwei Huang (Princeton) DSL Spectrum Management March 2007 9 / 26
-
8/6/2019 Dsl Management
19/42
Optimal Spectrum Balancing
Global optimization based on dual decomposition
Key: the duality gap is asymptotically zero under frequency-sharingproperty
R2
1R
1R
target
A
C
B
E L l
l
D
w = 0
w = 1
w =
w = +
XYX Y
c Cendrillon et. al., ICC, 2004
Jianwei Huang (Princeton) DSL Spectrum Management March 2007 10 / 26
l l
-
8/6/2019 Dsl Management
20/42
Optimal Spectrum Balancing
Partial Lagrangian:
L (p1 ,..., pN ) =n
w nk
log 1 + SINR k n n
nk
p k n P maxn
Decompose K nonconvex subproblems, one for each tone k :
maximize{p k n }
n 0 n
w n log 1 + SINR k n n
n p k n
Joint exhaustive search of optimal transmission power of all usersOptimal values of 1 ,..., N can be found using bisection orsubgradient search
Jianwei Huang (Princeton) DSL Spectrum Management March 2007 11 / 26
-
8/6/2019 Dsl Management
21/42
-
8/6/2019 Dsl Management
22/42
-
8/6/2019 Dsl Management
23/42
-
8/6/2019 Dsl Management
24/42
Iterative Water lling
-
8/6/2019 Dsl Management
25/42
Iterative Water-lling
ProsAutonomous: no explicit communication among users (interferenceplus noise can be locally measured)Low computational complexity of O (KN ): separable across users andtonesAchieve better performance than the current practice
ConsSelsh optimizationNo consideration for damages to other users
Highly suboptimal in the mixed CO/RT case
Jianwei Huang (Princeton) DSL Spectrum Management March 2007 14 / 26
Autonomous Spectrum Balancing
-
8/6/2019 Dsl Management
26/42
Autonomous Spectrum Balancing
Key idea: reference line - static pricing for static channelA virtual line representative of the typical victim in the networkGood choice: the longest CO lineParameters (power, noise, crosstalk) are publicly known
Each user will choose its transmit power to protect the reference line
Jianwei Huang (Princeton) DSL Spectrum Management March 2007 15 / 26
Reference Line
-
8/6/2019 Dsl Management
27/42
Reference Line
CP
RT
RT
RT
CP
CO CP
CP
Jianwei Huang (Princeton) DSL Spectrum Management March 2007 16 / 26
Reference Line
-
8/6/2019 Dsl Management
28/42
Reference Line
Actual Line
Reference Line
CO
CPCO
RT CP
RT
RT
CP
CP
CP
Jianwei Huang (Princeton) DSL Spectrum Management March 2007 16 / 26
Reference Lines Rate
-
8/6/2019 Dsl Management
29/42
Reference Line s Rate
User ns obtains the reference line parameters locally
Length & Location Reference Crosstalk:Reference Noise:
Reference Power:OperatorReference LineDatabase
pk,ref
k,ref
k,ref n
Jianwei Huang (Princeton) DSL Spectrum Management March 2007 17 / 26
Reference Lines Rate
-
8/6/2019 Dsl Management
30/42
Reference Line s Rate
User ns obtains the reference line parameters locally
Length & Location Reference Crosstalk:Reference Noise:
Reference Power:OperatorReference LineDatabase
pk,ref
k,ref
k,ref n
The reference line rate
R ref n =k
log 1 +p k , ref
k , ref n p k n + k , ref
Interference only depends on user ns transmit power p k nLocally computable without explicit message passing
Jianwei Huang (Princeton) DSL Spectrum Management March 2007 17 / 26
Frequency Selective Water-lling
-
8/6/2019 Dsl Management
31/42
Frequency Selective Water llingUnder high SNR approximation of the reference line
Bk n p n = w n n + k , ref n / k , ref 1{p k , ref > 0}
m = n
k n , m p k m k n
+
Reference line isnot active in high frequency tones
Special case: traditional water-lling (ignore k , ref n / k
, ref )
Jianwei Huang (Princeton) DSL Spectrum Management March 2007 18 / 26
Frequency Selective Water-lling
-
8/6/2019 Dsl Management
32/42
Frequency Selective Water llingUnder high SNR approximation of the reference line
Bk n p n = w n n + k , ref n / k , ref 1{p k , ref > 0}
m = n
k n , m p k m k n
+
Reference line isnot active in high frequency tones
Special case: traditional water-lling (ignore k , ref n / k
, ref )
Jianwei Huang (Princeton) DSL Spectrum Management March 2007 18 / 26
Power
Traditional WaterFilling
Frequency
Interference & Noise
Frequency Selective Water-lling
-
8/6/2019 Dsl Management
33/42
Frequency Selective Water llingUnder high SNR approximation of the reference line
Bk n p n = w n n + k , ref n / k , ref 1{p k , ref > 0}
m = n
k n , m p k m k n
+
Reference line isnot active in high frequency tones
Special case: traditional water-lling (ignore k , ref n / k
, ref )
Jianwei Huang (Princeton) DSL Spectrum Management March 2007 18 / 26
Power
Active Reference Line
FrequencySelective WaterFilling
Frequency
Interference & Noise
Convergence of ASB Algorithm
-
8/6/2019 Dsl Management
34/42
Convergence of ASB Algorithm
ASB Algorithm: users update their individual power allocationaccording to best responses either sequentially or in parallel
TheoremASB algorithm globally and geometrically converges to the unique N.E. if the crosstalk channel is small , i.e.,
maxn , m , k
k n , m