1 Need of Time-of-day Internet Access Management Peak-hour bandwidth utilization 100%(9 a.m.–3 a.m.) Peak-hour drop rate > 3 Mbps Peak-hour usage: Heavy.

1

Need of Time-of-day Internet Access Management

• Peak-hour bandwidth utilization 100%(9 a.m.–3 a.m.)

• Peak-hour drop rate > 3 Mbps• Peak-hour usage: Heavy : normal = 13.08: 1

User Group

Average Usage (Bytes)

2GHU 17,004,173

1GHU 24,530,722

100MHU 43,410,386

NU 4,754,360

LU 1,736,643

Q: What problems do we suffer over free-of-charge or flat rate network?- Ex: NTU dorm networks

How to 1) Manage the time-of-day Internet access2) Design an incentive control scheme

2

Research on Time-of-day Internet Access Management by

Quota-based Priority Control

Presented by Shao-I Chu Advisor: Dr. Shi-Chung Chang

Date: June 13, 2007

3

Outline

• Existing Quota-based Priority Control• User Behavior: Prudent and Myopic• Design of Management scheme I:

Game theoretic virtual pricing• Design of Management scheme II:

Heuristic-based Quota Scheduling• Performance Comparisons• Conclusions

4

Outline

• Existing Quota-based Priority Control

• User Behavior: Prudent and Myopic• Design of Management scheme I: Game

theoretic virtual pricing• Design of Management scheme II:


5

Quota-based Priority Control (QPC)

• Solve abusive and unfair usage• Missions of network manager

– Meet majority users’ basic demand – Limit heavy users’ abusive usage

• QPC Services– Regular service (high priority) – daily

quota limitation– Custody service (low priority)

6

Existing QPC Architecture

Dormitory Network

Dormitory Network

Router(Cabletron SSR2000)

TANet

Router(packet engine 5022)

Metering Router(Cisco 7513)

NTU Domain

54Mbps

QoS Router

DB

Accounting and Traffic Control Server

Web-based Service Management Server

Meter Reading Server

54Mbps

Merits of QPC - Daily congestion improved by 48%- Over 91% users’ usage encouragedWeakness of QPC- No consideration on temporal effect

7

Outline

• Existing Quota-based Priority Control• User Behavior: Prudent and Myopic• Design of Management scheme I: Game



8

Myopic and Prudent Behaviors under QPC

• Myopic User: no consideration on quota limitation (6:00 am quota renewal)

• Prudent User: careful allocation of one’s quota

0 2 4 6 8 10 12 14 16 18 20 22 240

1

2

3

4

5

6

7

8x 10

7

Day Time

Inte

rnet

Usage (

Byte

s)

0 2 4 6 8 10 12 14 16 18 20 22 240

0.5

1

1.5

2

2.5

3

3.5

4

4.5

5x 10

7

Day Time

Inte

rnet

Usage (

Byte

s)

9

How to Design the Management Schemes

M1) How to effectively manage the time-of-day Internet access by utilizing minimal empirical data

M2) How to design a simple and incentive control scheme for easy acceptance by users

M3) How to combine the existing QPC architecture

M4) How to construct a design methodology for a changing network

10

Ex: Time length: peak: off-peak=1:1

Design of Management Schemes

• Quota Scheduling– Different quota allocations for different time

periods

Peak Hours

Off-peak Hours

User B

• Virtual Pricing – price=number of quota per byte– Price varies with time

Peak Hours

Off-peak Hours

User A

Peak Hours

Off-peak Hours

User B

Peak Hours

Off-peak Hours

User A

Incentive!!

11

Contributions of This Thesis• Propose virtual pricing and compare it with

quota scheduling and for time-of-day Internet access management– effective by utilizing minimal empirical data to

model user behaviors– incentive and flexible– easily combined with QPC– generic design methodology constructed for a

changing network

12

Outline

• Existing Quota-based Priority Control• User Behavior: Prudent and Myopic• Design of Management scheme I:

Game theoretic virtual pricing• Design of Management scheme II:


13

Challenges for Virtual Pricing Design

P1) How to exploit empirical to model user response w.r.t. price

P2) How to design a pricing policy to maximize bandwidth utilization

P3) How to design a simple pricing policy for user acceptance

P4) How to exploit the existing hardware and software of the legacy network

P5) How to design a methodology

To answer P3) and P4) Static Time-of-day Pricing (TDP)

14

Design Methodology of TDPMethodology

Step 1: Pilot Experiment and Analysis

Baseline Experiment Analysis

QPC Experiment Analysis

Step 2: Empirical User Demand Model Construction

Step 3: Time-of-day Pricing Design Using Game Theoretic Problem

Formulation

Step 4: Network Performance and User Usage Prediction by Simulation

Manager΄s Decision: New Policy Needed?

Measurement Data

Pricing Policy

Control Flow

Data FlowYes

Managed Network

QoS & Metering Router

Internet Access


Web-based Service Management ServerD

BAccounting and

Traffic Control Server

Intranet Traffic

Network Performance Monitoring

No

Manager΄s Review and Adjustment

Network Performance

MonitoringManager’s Decision:

New Policy Needed?

To answer P5): How to design a methodology

15







Formulation



Measurement Data

Pricing Policy

Control Flow

Data FlowYes

Managed Network


Internet Access



BAccounting and


Intranet Traffic


No





Step 1):Baseline Experiment - No quota limitation - Characterize network problem and user original demandQPC Experiment

- Daily quota- Provide data for constructing user models

16







Formulation



Measurement Data

Pricing Policy

Control Flow

Data FlowYes

Managed Network


Internet Access



BAccounting and


Intranet Traffic


No


Step 2: Empirical UserDemand Model Construction

Step 2):Construct myopic and prudent User behavior - Varies with price profile and demand - User preference estimated by QPC experiment

17







Formulation



Measurement Data

Pricing Policy

Control Flow

Data FlowYes

Managed Network


Internet Access



BAccounting and


Intranet Traffic


No


Step 3: Time-of-day Pricing Design Using Game-theoretic Problem

Formulation

Step 3):Leader: network manager - Maximize the total bandwidth utilization - keep the total demand below the capacityFollowers: users - Maximize their own benefits

18







Formulation



Measurement Data

Pricing Policy

Control Flow

Data FlowYes

Managed Network


Internet Access



BAccounting and


Intranet Traffic


No


Step 4: Network Performance and User Prediction by Simulation

Step 4):Perform numerical assessment based on empirical data under pricing policy by step 3Exploit the experimental data of step 1 and user demand model constructed by step 2 to simulate user behavior.

19







Formulation



Measurement Data

Pricing Policy

Control Flow

Data FlowYes

Managed Network


Internet Access



BAccounting and


Intranet Traffic


No


Manager’s ReviewAnd Adjustment

20

Myopic and Prudent User Classification

price profile & daily demand

(baseline experiment)

Myopic User Prudent User

}max{/ kBi pQv }min{/ k

Bi pQv

B=Q/min{pk}A=Q/max{pk}0

(myopic)

1 (prudent)

Daily demand

To answer P1):How to exploit empirical data

to model user response w.r.t. price

21

Myopic User Model• Focus on short-term benefit maximization

• Maximize i’s own benefit at that time slot k only

kikkiki

vvpvU

ki,, )( Max

,

User preference

volume F(.

) S

atis

fact

ion

diminishing returns of scale

)()( ,,, kikikiki vFvU

22

Prudent User Model

• Focus on daily benefit maximization

• Maximize i’s total benefit from time slot k to time slot K

subject to the quota budget constraint

K

kttitti

ti

KktvvpvU

ti,,

},...,,{)( Max

,

,, ki

K

kttit Qvp

23

How to Estimate Individual User Preference

• Derive preference from optimal conditions

)(/ ,, kikki vFp

QPCkiki vvkiki vF,,,

|)(/1 ,,

User Usage Data under QPC

(pk=1)

24

Selection of F(.)

• Myopic user:

• Prudent user:

• User preference:

,*,

k

kiki p

v

Kktp

Qv

t

ti

K

ktji

titi ,...,, ,

,

,*,

QPCkiki v ,,,

Utility(Rate)=log(Rate) Utility(Volume)=log(Volume)

i.e., F(. )=log(. )

25

User Volume under Baseline Experiment

User Volume under QPC Experiment

{pk|k=1,2,…,K}

User Classification

Utility Function F(.)=log(. )

User Behavior Model w.r.t. Price

User Preferences

Myopic Prudent

To answer P1):How to exploit empirical data

to model user response w.r.t. price

26

TDP Design

• Manager’s Decision Problem:

Price Profile

Maximize total bandwidth utilization of regular service

Total User submission cannot exceed the bandwidth

Goal of Network Manager

When service is free or flat rate

To answer P2): How to design a pricing design

to maximize total bandwidth utilization

27

Leader-follower Model

Leader- Network ManagerGoal:

Maximize total bandwidth utilization

Follower- Users

Myopic UserMaximize short-term benefits

Prudent UserMaximize daily benefits

PriceVolume Volume

28

Analysis of TDP Policy

• Goals1) How the prices may induce user behavior an

d affect network performance

2) How TDP policy varies w.r.t user behavior

• Problem Settings- 3 users- 3 time units- bandwidth:10 units- price set

Preferences Time Slot 1 Time Slot 2 Time Slot 3

User 1 3 5 7

User 2 5 7 9

User 3 7 9 113,2,1 },5,4,3,2,1{ ipi

29

Why Needs User Differentiation

• Case I

– Pricing policy :prudent, Users: myopic • Case II

– Pricing policy : myopic, Users: prudent

Total Submitted Volume (Q=10)

Time Slot 1 Time Slot 2 Time Slot 3

Case I 15 21 13.5

Case II 3.49 3.33 13.03

Submitted volumes are not shaved (>10)

- Bandwidth utilization < 50% at time slots 1 and 2- Congestion happens at time slot 3

30

Pricing Policy for Prudent Users

• Hypotheses: – The higher user preference the higher price

for a time slot • Analyses:

– Due to link capacity constraint

BTp

Qv

I

i k

kiK

ktti

kiI

iki

1

,

,

,

1

*,

K

jji

kiki

I

i

kikik W

BT

QWp

1,

,,

1

,, where,

I

i

kikik BT

QWp

1

,,*Total Submitted Volume (Q=10)


QPC Scheme 6.97 10 13.03

TDP Scheme 6.97 10 6.51

• Q=10 P=(1,2,3)• Q=25 P=(2,3,4)

Congestion!

31

Pricing Policy for Myopic Users

• The property that no longer holds• Analysis:

– Due to link capacity constraint

*3

*2

*1 ppp

kAi

kik BTp ,

BTp

kAi k

ki

,

kAi

kik BTp ,*

Total Submitted Volume (Q=10)


QPC Scheme 15 21 27

TDP Scheme 7.5 7 7

• Q=10 P=(2,3,1)• Q=20 P=(2,3,3)

Congestion! Not shaved!

32

Effectiveness Evaluation• Parameter Setting

– Peak hour 9 a.m. to 3 a.m – Quota replenishment point 6:00 a.m.

– Length of each time slot 10 minutes.

– Bottleneck bandwidth 54Mbps. – Admissible price set (per byte):

Ω={1, 1.1, 1.2, 1.3, 1.4, 1.5}

– Quota budget of each user 1G

Hypotheses1)Optimal price: 2)Drop rate of regular service 0 3)Peak-hour usage:

- Total submitted volume of regular service ↓

- User transmitted volume of Internet access ↓

**peakpeakoff pp

33

Peak Shaving and Load Balancing Effects

• Optimal Price 3.1,1.1 ** peakpeakoff pp

0 2 4 6 8 10 12 14 16 18 20 22 240

10

20

30

40

50

60

Day Time

Mbps

Total Submitted Volume under QPC/TDP

Total Submitted Volume under QPC

Drop Rate under QPC/TDPDrop Rate under QPC

Available Bandwidth

Total submitted rate reduced by 11.53% during peak hoursDifference between peak and off-peak hours reduced by 31.21%

Peak-hour drop rate reduced to 0

TDP effectively manages the time-of-day Internet access !

34

Peak-hour Abuse Improvement• Abuse Index – Top 5 user Internet usage

QPC Scheme 500MB*2 QPC Scheme

Peak -hours -17.62% -4.4%

0 2 4 6 8 10 12 14 16 18 20 22 241

2

3

4

5

6

7x 10

8

Day Time

Top 5

User

Usage o

f In

tern

et

Access (

Byte

s)

QPC/TDP Scheme

QPC Scheme

35

Peak-hour Fairness Improvement• Fairness Index– Standard deviation of Internet

usage

0 2 4 6 8 10 12 14 16 18 20 22 241

1.5

2

2.5

3

3.5

4

4.5

5

5.5

6x 10

6

Day Time

Sta

ndard

Devia

tion o

f U

ser

Inte

rnet-

Access V

olu

me (

Byte

s)

QPC/TDP Scheme

QPC Scheme

QPC Scheme 500MB*2 QPC Scheme

Peak –hour -17.64% -8%TDP improves peak-hour

abuse and unfairness

36

Policy Adaptation to Changes

Short time period for data collection:– Baseline and QPC experiments will be

conducted for a short period (1 week each) – Only conducted at the beginning of a new

academic year• Fast policy design and evaluation

– Takes several minutes in the case of the NTU dormitory network with 5000+ users

37

Outline




38

Load Balancing-based Quota Scheduling (LB-QS)

• Objective:– Equalize the average traffic of peak and off p

eak hours

• Designed Quota Scheduling:

peakoff

peakoff

peak

peak

T

IQ

T

IQ

QTT

TQ

peakoffpeak

peakpeak

Q

TT

TQ

peakoffpeak

peakoffpeakoff

39

Peak Shaving-based Quota Scheduling (PS-QS)

Off-peak Hours Peak Hours

Total Submission Rate under QPC

Bandwidth Limitation

Estimated User Quota Usage

User Quota Usage Total Submission

d

d

Scheduled Quota

40

Outline




41

Comparisons of TDP and QS: Load Balancing & Peak Shaving

• Peak Shaving Index (PSI): – Average total submission rate of peak hours

• Load Balancing Index (LBI)– Difference of average total submission rates

between peak and off-peak hours

LB-QS PS-QS TDP

LBI (Mbps) 16 9.98 9.28

PSI (Mbps) 57.42 54.75 51.35

• LBI and PSI under PS-QS improved by 37.6% and 4.7% over LB-QS because of considerations on user preferences over time

Evaluated over the empirical data of NTU dormitory network

LB-QS (Qpeak,Qoff-peak)=(750MB, 250MB) No user usage data needed

PS-QS (Qpeak,Qoff-peak)=(620MB, 380MB) User usage data of QPC

TDP (Ppeak,Poff-peak)=(1.3, 1.1) User usage data of QPC Baseline data (no control)

42

Comparisons of TDP and QS: Total Submission Rate

0 2 4 6 8 10 12 14 16 18 20 22 240

10

20

30

40

50

60

70

80

Day Time

Tot

al S

ubm

issi

on R

ate

of R

egul

ar S

ervi

ce (

Mbp

s)

TDP

PS-QSLB-QS

LINK CAPACITY

•a spike (congestion) at 9 a.m. because of no price and no user differentiation•PS-QS encourages more usage than LB-QS because of user preference

43

Comparisons of TDP and QS: Abuse and Fairness Improvement• Abuse Index (AI)

– Internet access volume by top 5 users • Fairness Index (FI)

– Standard deviation among all users’ usage

• TDP outperforms QS by at least 14%• PS-QS is better than LB-QS (user preferences )

LB-QS PS-QS TDP

AI (bytes) 226964566 206758615 173572422

FI (bytes) 2631251 2440906 2107364

44

Design Related Issues

LB-QS PS-QS TDP

Measurement Requirement

No user data

User data of QPC

User data of QPC and baseline

Calculation Complexity

Simple Simple Solve a leader-follower game

Implementation

Requirement

QS module at accounting and traffic

control server

Pricing module at accounting and traffic

control serverApplicability to

Traffic Pattern

If the peak hours are not contiguous but scattered over all time slots congestion at th

e quota renewal time TDP

0 2 4 6 8 10 12 14 16 18 20 22 240

10

20

30

40

50

60

70

80

Day Time

Tot

al S

ubm

issi

on R

ate

of R

egul

ar S

ervi

ce (

Mbp

s)

TDP

PS-QSLB-QS

LINK CAPACITY

45

Outline




46

Conclusions (1/2)

• Propose a incentive and simple control scheme TDP over free-of-charge or flat rate network (M2,P3)

• TDP is easily implemented over QPC (M3, P4)

• TDP develop empirical data-based user model (P1)– Myopic and prudent users

• TDP uses game-theoretic design to maximize bandwidth utilization (P2)– Network manager as leader, users as followers

47

Conclusions (2/2)• TDP effectively manages the time-of-day

Internet access traffic (M1)– Peak-hour abuse and fairness improved by

14% above over QS– Load balancing and peak shaving reduced by

24% and 9%

• Generic methodology of TDP is proposed for a changing network (M4, P5)– Two short-period data collections– Fast evaluation and design in several minutes– Apply to campus, government, community and

corporate LANs

48

Thanks for your attention!

1 Need of Time-of-day Internet Access Management Peak-hour bandwidth utilization 100%(9 a.m.–3 a.m.) Peak-hour drop rate > 3 Mbps Peak-hour usage: Heavy.

Documents

day pricing design

methodology slide

qpc myopic user

design methodology of

quota renewal prudent

myopic user prudent

qpc experiment slide

prudent user behavior