Slide

Self-Organized Resource
Allocation in LTE Systems with
Weighted Proportional Fairness
I-Hong Hou and Chung Shue Chen
Motivation
• 4G LTE networks are being deployed
• With the exponentially increasing
number of devices and traffic,
centralized control and resource
management becomes too costly
• A protocol for self-organizing LTE
systems is needed
Challenges
• LTE employs OFDMA
• Link gains can vary from subcarriers to
subcarriers due to frequency-selective
fading
• Need to consider interference between
links
• A protocol needs to achieve both high
performance and fairness
Our Contributions
• Propose a model that considers all the
challenges in self-organizing LTE
networks
• Identify three important components
• Propose solutions for these components
that aim to achieve weighted
proportional fairness
Outline
•
•
•
•
•
•
System Model and Problem Formulation
An algorithm for Packet Scheduling
A Heuristic for Power Control
A Selfish Strategy for Client Association
Simulation Results
Conclusion
System Model
• A system with a number
of base stations and
mobile clients that
operate in a number of
resource blocks
• A typical LTE system
consists of about 1000
resource blocks
• Each client is associated
with one base station
Channel Model
• Gi,m,z := the channel gain
between client i and base
station m on resource
block z
• Gi,m,z varies with z, so
frequency-selective
fading is considered
Channel Model
Interference
Signal
• Suppose base station m
allocates Pm,z power on
resource block z
• Received power at i is
Gi,m,zPm,z
• The power can be either
signal or interference
• SINR of i on z can be
hence computed as
Pi,m,zGi,m,z
SINRi,z =
N i,z + å Pl,zGi,l,z
l¹m
Channel Model
• Hi,m,z := data rate of i
when m serves it on z
• Hi,m,z depends on SINR
• Base station m can serve
i on any number of
resource blocks
• øi,m,z := proportion of time
that m serves i on z
• Throughput of i:
ri = å Fi,m,z H i,m,z
z
Problem Formulation
• Goal: Achieve weighted proportional
fairness
• Max
wi logri (wi := weight of client)
å
i
• Choose suitable øi,m,z (Scheduling)
• Choose Pm,z (Power Control)
• Each client is associated with one base
station (Client Association)
An Online Algorithm for
Scheduling
• Let ri[t] be the actual throughput of i up to
time t
• Algorithm: at each time t, each base
station m schedules i that maximizes
wiHi,m,z/ri[t] on resource block z
• Base stations only need to know
information on its clients
• The algorithm is fully distributed and can
be easily implemented
Optimality of Scheduling
Algorithm
• Theorem: Fix Power Control and Client
Association,
limt®¥ å wi logri [t] = max å wi logri
i
i
• The scheduling algorithm optimally solves
Scheduling Problem
• Can be extended to consider fast-fading
channels
Challenges for Power Control
• Find Pm,z that maximizes
å w logr
i
i
i
• Challenges:
• The problem is non-convex
• Need to consider the channel gains
between all base stations and all clients
• Need to consider the influence on
Scheduling Problem
Relax Conditions
• Assume:
• The channel gains between a base station
m and all its clients are the same, Gm
• The channel gains between a base station
m and all clients of base station o are the
same Gm,o
• We can directly obtain the solutions of
Scheduling Problem
A Heuristic for Power Control
• Propose a gradient-based heuristic
• The heuristic converges to a local optimal
solution
• The heuristic only requires base stations to
know local information that is readily
available in LTE standards
• Can be easily implemented
Client Association Problem
• Assume that each client aims to choose
the base station that offers most
throughput
• Consistent with client’s own interest
• In a dense network, a client’s decision has
little effects to the overall performance of
other clients
Estimating Throughput
• To know the throughput that a base station
offers, client needs to know:
• Hi,m,z : throughput on each resource block,
can be obtained by measurements
• øi,m,z : amount of time client is scheduled
• Develop an efficient algorithm that
estimates øi,m,z
• Solves Client Association Problem
Simulation Topology
X25
X16
X16
X9
500 m
Simulation Settings
• Channel gains depend on:
• Distance
• Log-normal shadowing on each frequency
• Rayleigh fast fading
Compared Policies
• Default
– Round-robin for Scheduling
– Use the same power on all resource blocks
– Associate with the closest base station
• Fast Feedback: has instant knowledge of
channels
• Slow Feedback: only has knowledge on
time-average channel qualities
Simulation Results
Total Throughput (Mbps)
300
250
200
150
100
50
0
Default
Slow Feedback
Fast Feedback
Simulation Results
1
0.9
0.8
0.7
CDF
0.6
0.5
Default
0.4
0.3
Slow Feedback
0.2
Fast Feedback
0.1
0
0
2
4
6
Throughput (Mbps)
8
10
Conclusion
• We investigate the problem of selforganizing LTE networks
• We identify that there are three important
components: Scheduling, Power Control,
Client Association
• We provide solutions for these problems
• Simulations show that our protocol
provides significant improvement over
current Default policy