femto-matching: efficient traffic offloading in heterogeneous ... · femto-matching: efficient...
TRANSCRIPT
Femto-Matching: Efficient Traffic Offloadingin Heterogeneous Cellular Networks
Wei Wang , Xiaobing Wu, Lei Xie and Sanglu Lu
Nanjing University
April 28, 2015
1/1
Heterogeneous Cellular Networks
Micro-cell
femto-cell
femto-cell
femto-cell
femto-cell
Micro-cell
femto-cell
Marco-cell
• Cellular networks use multiple layers of basestations to im-prove spatial utility
• Small cells such as femtocell or WiFi APs help offloadingthe traffic form the macro-cell
2/1
User Association Problem
• Mobile devices can connect tomultiple basestations
• Basestations provide differentservice qualities
• Traditional approach- strongest signal- lowest price- highest transmission rate
3/1
Problems for Existing Approaches
Scenarios of suboptimal solutions
• For the user- overly crowded basestations
• For the carrier- low utilization of femto-cells- overload of marco-cells
femto A
femto B
1M
1M
1M
1M
4 M
4/1
Problems for Existing Approaches
Scenarios of suboptimal solutions
• For the user- overly crowded basestations
• For the carrier- low utilization of femto-cells- overload of marco-cells
femto A
femto B
1.3M
2M
1!"M
1.3
M
2 M
4/1
Problems for Existing Approaches
Scenarios of suboptimal solutions
• For the user- overly crowded basestations
• For the carrier- low utilization of femto-cells- overload of marco-cells
femto A
femto B
1M
1.3 M
1M
1M
1.3 M
Cannot connect
to BS A due to
resource
constraints
1M
Associate to
macro-cell
1.3 M
4/1
Problems for Existing Approaches
Scenarios of suboptimal solutions
• For the user- overly crowded basestations
• For the carrier- low utilization of femto-cells- overload of marco-cells
femto A
femto B
1M
1 M
1M
1M
1 M
Cannot connect
to BS A due to
resource
constraints
1M
Associate to
macro-cell
1M
1M
4/1
Randomly Deployed Network
0 10 20 30 40 50 60 700
10
20
30
40
50
60
70
meters
mete
rs
An example for randomly deployed network
5/1
Randomly Deployed Network
0 10 20 30 40 40 60 700
10
20
30
40
50
60
70
meters
me
ters
“orphan nodes” 20%→ 5%
5/1
Key Challenges
• Misaligned objectives for users and network operators- Users want better throughput- Operators want better resource utilization
• Both the femto-cells and the users are randomly distributed• Require a global view to fully optimize the system• Mobile devices move all the time• One device usually only associates with one basestation
6/1
Problem Formulation
• Objective: maximize the overall utility- Proportional Fairness
max∑
i
log(∑
j
cij rijaij
)• Constraints:
- Basestations split their resources to associated users- Users only associate to one basestation∑
i cij ≤ 1 ∀j ∈ B,∑j aij ≤ 1 ∀i ∈ U ,
aij ∈ {0,1}, cij ≥ 0 ∀i ∈ U , j ∈ B.
7/1
Problem Transformation
• Mixed integer programming problem, but can be solved op-timally
• Key observations• Within a single cell, the resources will be divided evenly in
the proportional fairness case:
e.g., k − 1 users in BS j, with rate r1j , r2j , . . .
Each user takes 1k−1 of resources (time slots, RB,..), through-
put r1jk−1 , r2j
k−1 , . . .
Overall utility:
k−1∑i=1
logrij
k − 1=
k−1∑i=1
log rij − (k − 1) log(k − 1)
8/1
Problem Transformation
• A new user k with rate rkj joins
• Rate of existing users reduces to rijk
• Overall utility:
k−1∑i=1
logrij
k+ log
rkj
k=
k−1∑i=1
log rij + log rkj − k log k
• Marginal utility of user k :
log rkj + (k − 1) log(k − 1)− k log k
9/1
Problem Transformation
Converting the problem to an equivalent maximum weighted match-ing problem
Original network
Users
Base Stations
U1 U2 U3 U4
BS1 BS2
r r r rr11 21 31 32 42
• Splitting BS to virtualBSs
• Weights are marginalutilities
log (r /4)log (r /4)
u u u u
v v
log r11
v v v
1 2 3 4
1 1 1 2 21 2 13 2
log r log r31log r42
42
log (4r /27)11
log r 32
32log (r /4)11 log (r /4)21 log (r /4)31
log (4r /27)21 log (4r /27)31
21
10/1
Distributed Solution
• Exploit the special structure of the problem to design a dis-tributed matching algorithm
• Basic idea- Divide the edge weight to two parts, maintained separately
by the BS and user- Use the price to characterize the importance of the resource
• Auction process- Initialization: BSs set the initial prices for all virtual BSs;
Users estimate transmission rates- Iterative auction
* BSs announce the lowest price among virtual BSs* Users submit bids to the BS with highest gain* BSs select the user with highest bid and adjusts prices
• Finalize the association
11/1
Handling Mobility
• Incremental adjustment
• Node join- New node bids for the available resources in femtocells- Considering the cascading re-association
• Node leave- BS reduces the prices of the vacancy- Restart auction only when necessary
12/1
Performance Analysis
• Performance metrics- Offloading efficiency η:
ratio of users served by femtocells, reflects the efficiency offemtocells
• Both the femtocells and users are distributed as PoissonPoint Process, system parameters:
- l : Load factor l = λu/λf- κ: Number of users can be served by one femtocell
• Consider the efficiency of different schemes:- Associate to nearest BS- Matching schemes
13/1
Performance Comparation
• Associate to nearest BSOffloading efficiency around 74% due to randomness in deployment
l κ = 1 κ = 2 κ = 3 κ = 4 κ = 5 κ = 61 0.5851 0.8474 0.9483 0.9835 0.9950 0.99852 0.6636 0.8230 0.9110 0.9568 0.97963 0.6980 0.8132 0.8877 0.93414 0.7176 0.8080 0.87215 0.7303 0.80486 0.7393
• Matching schemeOffloading efficiency approaching 1, under higher network density.
14/1
Experiment Setup
• Experiments setups- Simulations with randomly generated networks- Trace-driven simulations on UIUC UIM trace
• Algorithms- Associate to nearest- RAT selection game- College admission algorithm- Femto-Matching
15/1
Experimental Results I
50 100 1500
0.05
0.1
0.15
0.2
0.25
0.3
N
1−
η
Nearest
College
RAT game
FemtoMatching
lower bound
Femto-matching has lowest ratio of not offloaded users
16/1
Experimental Results II
50 100 1500.6
0.7
0.8
0.9
1
N
Fa
irn
es
s i
nd
ex
College
RAT game
FemtoMatching
Femto-matching provides better fairness
17/1
Experimental Results III
1 2 3 4 5 6 7 80
20
40
60
Load of femtocellsNu
mb
er
of
fem
tocells
College
RAT game
FemtoMatching
50 100 150
400
600
800
N
Nu
mb
er
of
rou
nd
s
Simulation result
Curve fitting
Femto-matching provides better load balancing amongfemtocells and has low computational cost
18/1
Conclusion and Future Works
• Matching provides a good way to smooth out the random-ness in deployment
• It is possible to distributively calculate the optimal propor-tional fairness allocation
• Future researches• Detailed performance evaluation for mobility• Truthfulness in auction
19/1
Q & A
Thanks!
20/1