Towards Viable Large Scale Heterogeneous Wireless Networks

Towards Viable Large Scale Heterogeneous Wireless Networks

Ph.D. Dissertation Defenseby

Rahul AminElectrical and Computer Engineering

Research Advisor: Dr. Jim Martin

Committee Members:Dr. Harlan B. RussellDr. Daniel Noneaker

Dr. Brian DeanDr. Melissa Smith

2

Outline

• Introduction

• Motivation and Background

• Thesis Statement

• System Model

• Research Phase I Summary – Resource Allocation Problem

• Research Phase II Study – Practical Implementation Issues

• Conclusions

3

Introduction• The demand for wireless data traffic is outgrowing current

supply capabilities

– Proliferation of mobile devices in the last decade has created an exponential growth in traffic demand

– FCC projects a 275 MHz spectrum deficit by 2014 if no new spectrum is made available for broadband usage

– Utilizing available spectrum in the most efficient manner becomes paramount

Source: Opennetsummit April 2012

4

Introduction (Contd.)

• While there are areas of spectrum that are over-utilized, there are other areas of spectrum that are underutilized

• This has renewed interest in techniques that attempt to improve spectral efficiency through co-operation at the radio/network level – Bottom-up Approaches

• Examples: Cognitive networks, Dynamic Spectrum Access networks

– Top-down Approaches• Examples: Heterogeneous wireless networks, Wi-Fi Offloading, Femtocells/Picocells

• Heterogeneous wireless networks (hetnets) made up of several cellular (LTE, WiMAX, HSPA) and WLAN (Wi-Fi) Radio Access Technologies (RATs) is the focus of our study

5

Motivation and Background

• Due to the widespread deployment of several wireless RATs, it is quite common for any geographical location to be covered by more than one wireless network

• Current practices lead to sub-optimal spectrum usage– Wireless networks are built independently– Individual networks attempt to achieve best performance within its own

network, generally ignoring impact of co-located networks– Users are required to select the active access network

• A fundamental motivation for our research is that enhancing access and use of spectrum requires a combination of cognitive device capabilities AND a component of resource allocation that operates at the global level [VT ’11]

Source: [VT ’11] J. Martin, R. Amin, A. Eltawil, A. Hussien, “Limitations of 4G Wireless Systems,” Proceedings of Virginia Tech Wireless Symposium, June 2011.

6

Motivation and Background (Contd.)

• Frameworks have been defined by both IEEE and 3GPP groups to allow a hierarchical/centralized control of resources managed by multiple RATs

– IEEE Heterogeneous Wireless Frameworks:• 802.21 – Seamless mobility through networks based on different radio

access technologies• 1900.4 – Co-ordinated network-device decision making to aid in the

optimization of radio resource usage, including spectrum access control

– 3GPP Heterogeneous Wireless Frameworks:• Common Radio Resource Management, Joint Radio Resource Management,

Multi-access Radio Resource Management– Local resource managers of different wireless technologies interact with a

centralized entity to jointly optimize the process of resource allocation

7

Motivation and Background (Contd.)

• From a user device perspective, numerous reconfigurable architectures have been proposed that consist of flexible computing structure that can be reconfigured to connect to various RATs– Energy-efficient reconfigurable device architectures (MPSoC) are being

investigated based on various hardware components such as ASICs, FPGAs and DSPs [ICCD ’11]

• A common design metric among all platforms is reducing energy consumption that restricts both the capabilities of the device and the design choices that are available

• Enough progress has been made at both the system architecture level and at the user device level to make the implementation of a real hetnet system feasible in the near future

Source: [ICCD ’11] A. Hussien, A. Eltawil, R. Amin, J. Martin, “Energy Aware Task Mapping Algorithm for Heterogeneous MPSoC based Architectures,” Poster Proceedings of IEEE International Conference on Computer Design, October 2011.

8

Thesis Statement

• Our research addresses the resource allocation problem and practical implementation issues related to the creation of a hetnet system, where multiple RATs collectively provide a unified wireless network to a diverse set of users through co-ordination managed by a centralized Global Resource Controller (GRC) to improve the efficiency with which spectrum is utilized

– Characterize network efficiency in terms of four conflicting objectives: (i) spectral efficiency (ii) instantaneous fairness (iii) long-term fairness (iv) overall energy consumption

– Analyze spectral efficiency and energy consumption trade-offs involved with different user device choices and operating assumptions (Phase I – Study 1)

– Analyze achievable trade-offs between all four conflicting network performance objectives from a scheduling perspective (Phase I – Study 2)

– Analyze performance gains of a centralized solution compared to a distributed solution while taking RAT-specific implementation details and centralized control overhead into account (Phase II Study)

9

System Model

AWS #1

Global Resource Controller (GRC)

cUE

HeterogeneousWirelessNetwork

Local Resource Controller (LRC)

Radio Link 1

Radio Link i

Radio Link Aggregation

TCP/IP

Internet

Carrier’s Backend

WVLINK Ingress/Egres

s

Location Services

BatteryController

LocationController

Bandwidth Controller

Radio Link 1

Radio Link i

Filters, modems, coding logic

FPGA resources

AWS #2 AWS #i LTEeNodeB

LTEeNodeB

LTEeNodeB

WiMaxBS

WiMAXBSWiMAX

BS

S1-MME

S101

MME PDG

S1-U

S11

Mapping to 3GPP

10

Network Efficiency Metrics

• Spectral Efficiency– represented as the ratio of (data) rate allocated to each user

in the system to the total spectrum used:

• Long-Term Fairness– relates to the difference in rates allocated to each user over

long time-scales– computed using Jain’s Fairness Index as follows:

11

Network Efficiency Metrics (Contd.)• Instantaneous Fairness– If support for real-time traffic is assumed, this metric is

computed using the proportion of users whose minimum data rate requirements per scheduling interval are satisfied:

– If support for only best-effort traffic is assumed, this metric is computed using Jain’s Fairness Index for each scheduling interval:

12

Network Efficiency Metrics (Contd.)• Overall Energy Consumption– Hardware-based model: Assumes fixed energy consumption as

long as a radio is connected to a RAT

– Data Transfer-based model: Assumes radio can be operated in ‘deep sleep’ mode, where there is no energy consumption, when there is no active transmission/reception

13

Research Phase I Summary -Resource Allocation Problem

14

Research Overview

• Performed using high-level system modeling approach via MATLAB/AMPL simulations

• Heuristic Algorithm Simulation Study – Analyze achievable tradeoffs in terms of network efficiency measures of spectral efficiency

and energy consumption due to the benefits of network co-operation based on different user device assumptions, network topologies and network outages

– Algorithm accounts for spectral efficiency, instantaneous fairness and long-term fairness measures by following a two-step scheduling approach

• Optimization-based Algorithm Simulation Study– Analyze achievable tradeoffs in terms of all four network efficiency measures using a utility

function-based and a weighted sum approach for a limited set of user device assumptions, network topologies and network outages

– Algorithm accounts for instantaneous fairness using admission control procedure; the weights of the other three attributes computed using Analytic Hierarchy Process

15

Results Summary

• A linear increase in spectral efficiency due to benefits of network co-operation (multi-access network diversity) has an order of magnitude higher increase in energy consumption based on current FPGA-based reconfigurable hardware [ICCCN ’11]

• Network topology and various user device hardware assumptions (ASIC, FPGA) have a significant impact on the spectral efficiency and energy consumption tradeoffs [WCNC ’12] – Exponential tradeoff for balanced network topology– Linear tradeoff for unbalanced network topology– ASIC-based hardware reduces energy consumption increase by almost 5x

• We show an increase of up to 56.7% in the multi-attribute utility measure for our algorithm compared to other widely used algorithms such as Max-Sum Rate, Max-Min Fairness, Proportional Fairness and Min Power [JSAC ’13]

Source: [ICCCN ’11] J. Martin, R. Amin, A. Eltawil, A. Hussien, “Using Reconfigurable Devices to Maximize Spectral Efficiency in Future Heterogeneous Wireless Systems,” Proceedings of IEEE International Conference on Computer Communications and Networks, Aug 2011.[WCNC ’12] R. Amin, J. Martin, A. Eltawil, A. Hussien, “Spectral Efficiency and Energy Consumption Tradeoffs for Reconfigurable Devices in Heterogeneous Wireless Systems,” Proceedings of IEEE Wireless Communications and Networking Conference, April 2012.[JSAC ’13] R. Amin, J. Martin, J. Deaton, L. DaSilva, A. Hussien, A. Eltawil, “Balancing Spectral Efficiency, Energy Consumption, and Fairness in Future Heterogeneous Wireless Systems with Reconfigurable Devices,” IEEE Journal on Selected Areas in Communications, May 2013.

16

Research Phase II Study -Practical Implementation Issues

17

Research Overview

• Performed using detailed protocol level simulator, ns-2, to model each RAT and the management overhead required for a hetnet system

• Greedy Sort-based Algorithm Simulation Study – Analyze the performance gains for a centralized GRC solution

compared to a distributed solution using our four network efficiency metrics

– Identify technical challenges associated with the management/operation of a hetnet system

– Analyze the overhead caused by the centralized GRC solution relative to the overall system (data) throughput

18

Simulation Description

• 2 WiMAX (802.16e) BSs, 6 Wi-Fi (802.11g) APs in 2 * 2 km2 area• 100 users

- Each user can use any available RAT- Three user movement patterns: (i) Linear Movement (ii) Random Waypoint (Speed: 2 mph) (iii) Random Waypoint (Speed: [2, 20 mph])

19

System Assumptions

• Each RAT (Wi-Fi and WiMAX) uses an adaptive Modulation and Coding Scheme (MCS)

• Wi-Fi MAC independently achieves max-min fairness scheduling objective via DCF which employs CSMA/CA with binary exponential backoff algorithm

• WiMAX MAC independently achieves proportional fairness scheduling objective via Deficit Weighted Round Robin scheduler

• The GRC operates on a five-second scheduling interval

• Each user device has an ASIC-based Wi-Fi and WiMAX radio; only one radio can be used for a data connection at any given time (integral association)

• Infinitely backlogged downlink data traffic for each user device (Constant Bit Rate traffic sent at 25 Mbps over TCP transport layer)

• Network re-association management implemented using IEEE 802.21’s Media Independent Handover Function (MIHF)

20

Extended System Model

• MIHF implemented at the 2.5 Layer of the OSI stack• cUE periodically sends link parameter report to the GRC• Using link parameter report, GRC periodically computes cUE-to-RAT

association mapping• Link Parameter Report Generation: (i) Periodic scanning (5-second basis) (ii) Location-based

21

Greedy Sort-based Algorithms• For each RAT, sort each user in descending order based on the user’s maximum achievable

data rate via the corresponding RAT– In case of ties, put user with lowest achievable data rate via all connectivity options first

• Greedily associate best unassociated user to a RAT in each association decision round based on two metrics: achievable total system throughput and lowest user throughput

• Use scheduling objective (Proportional Fairness or Max-Min Fairness) of each RAT in each round to determine achievable total system throughput and lowest user throughput metrics via following two propositions:

22

Greedy Sort-based Algorithms (Contd.)

• Max Throughput Algorithm– Make decisions based on maximum achievable total system throughput

metric– In case of ties, use maximum lowest user throughput metric

• Max Fairness Algorithm– Make decisions based on maximum lowest user throughput– In case of ties, use maximum achievable total system throughput metric

• Each of the two algorithms terminates when each user is assigned to a RAT

23

Example (Max Throughput Algorithm)

a b

1

2

3

4

12

6

6

4

4

3

2

1

BS aMax-Min

Fair

BS bProportional

Fair

User 1 User 1

User 3 User 2

User 2 User 3

User 4 User 4

Sorted users for each BS

Round #

Total System Throughput (Next assoc.

BS a)

Lowest User Throughput through BS a

Total System Throughput (Next assoc.

BS b)

Lowest User Throughput

through BS b

Association Decision

1 12 12 4 4 User 1 – BS a

2 8 4 15 3 User 2 – BS b

3 11 4 14.5 1 User 3 – BS b

4 8.5 3 14 0.333 User 4 – BS b

Greedy Algorithm

Achievable Data Rate

BS

User

24

Example (Max Fairness Algorithm)

a b

1

2

3

4

12

6

6

4

4

3

2

1

BS aMax-Min

Fair

BS bProportional

Fair

User 1 User 1

User 3 User 2

User 2 User 3

User 4 User 4

Sorted users for each BS

Round #

Total System Throughput (Next assoc.

BS a)

Lowest User Throughput through BS a

Total System Throughput (Next assoc.

BS b)

Lowest User Throughput

through BS b

Association Decision

1 12 12 4 4 User 1 – BS a

2 8 4 15 3 User 3 – BS a

3 7.2 2.4 11 3 User 2 – BS b

4 9 2 10 2 User 4 – BS b

Greedy Algorithm

25

Spectral Efficiency Results

• Both centralized algorithms (location based) outperform distributed solution- Up to 99.2% increase in spectral efficiency (Linear movement pattern)

• Scan based solutions lead to unpredictable performance due to disruption of active TCP data connections

26

Fairness Results

• Both instantaneous and long-term fairness metrics for Centralized Max Fairness algorithm improve compared to distributed solution

- Up to 28.5% increase in instantaneous fairness metric (Random Waypoint Same Speed)

• Both metrics deteriorate for Max Throughput algorithm compared to distributed solution

27

Overall Energy Consumption Results

• Based on two components: (i) energy consumption per bit transmitted/received and (ii) the number of handovers

- Dominated by the first component as handover costs of low energy consuming ASIC-based hardware assumed

28

Energy Consumption Due to Handovers

• More accurate representation of increase in energy consumption that implements periodic association computations via the use of a centralized controller

• Number of horizontal handovers (WiMAX-to-WiMAX) lead to a drastic increase in energy consumption for centralized solutions

- Up to 794% increase (Linear movement pattern)

29

Spectral Efficiency vs. Energy Consumption Trade-off

• Increase in energy consumption due to handovers outgrows increase in spectral efficiency by a factor of approximately 5 to 14 for various movement patterns!!!

30

Overhead Results

• The overhead required by centralized solution based on IEEE 802.21 framework is very manageable

- Lower than 4.7% of overall throughput for random movement patterns- Up to 18.3% of overall throughput for linear topology which does not use

resources of all available RATs (rare case)

31

Conclusions

• Analyzed achievable tradeoffs in terms of network efficiency measures of spectral efficiency and energy consumption due to the benefits of network co-operation based on different user device assumptions, network topologies and network outages

• Analyzed achievable tradeoffs in terms of all four network efficiency measures using a utility function-based and a weighted sum approach for a limited set of user device assumptions, network topologies and network outages

• Analyzed performance gains for a centralized GRC solution compared to a distributed solution using a detailed protocol-level simulator and IEEE 802.21 standards-based solution

• General trend in each study showed a higher increase in energy consumption compared to the increase in spectral efficiency- Advanced power management schemes for user devices operating in a hetnet system required- Resource allocation algorithms implemented at the GRC need to account for energy

consumption increase due to frequent re-associations

32

Thank You!