Network Management and Internet Pricing

S. Raghavan

J. Bailey

The Robert H. Smith School of Business

University of Maryland

Outline

• Goals• Ex-Post Charging Framework• Traffic Collection & Measurement• Models & Experiments

– Effective bandwidth model

– Moment-based model

• Ex-Post Charging Diffusion• Conclusions

Goals

• Develop a better mechanism for giving users and network administrators feedback regarding congestion costs (imposed on other users)

• Co-locate the economic incentives with the managers of the network

• Sacrifice economic optimality to achieve a pragmatic implementation

• Subscribe to the design philosophy of the Internet of heterogeneity and interoperability

Internet Pricing Goals

Ex-Post Charging

Network Managementi

Network externality

Network Managementj

Implementationj

Quality of Servicei

Extensibilityi

Scalabilityi

Securityi

Implementationi

Accountabilityi

Accountabilityj

Ex-Post Charging Concept

• Charging algorithm is known ex-ante, only utilization and burstiness metrics are unknown

• A portion of the charge may be determined ex-ante.

• ISP meters the link to collect aggregate traffic statistics.

• ISP charges the user at the end of the billing period using the charging algorithm and the utilization and burstiness statistics.

Features of Ex-Post Charging

• Simple to explain/understand– To ensure feasible implementation– For use in Service Level Agreements

• More accurate charging of congestion costs– Utilization prices have been assessed before– Addition of burstiness parameter

• Low Burden on System– Bayesian Updating

• No need for storage of gigabytes of trace data• Prevents potential security risk

– Non-intrusive mechanism (to collect data and bill).

Traffic Measurement• To fully develop a pricing mechanism, it is

essential to understand traffic measurement and collection.

• Achievable granularity with off the shelf technologies (low cost) – around 1ms.

• Finer granularity measurement can be achieved by using GPS synchronized clock and more expensive computing power. Traces used and analyzed– NLANR, New Zealand, University of Maryland (Van

Munching Hall)

Schema of Billing/Ex-Post Charging System

COMPUTATIO

N

MODULE

BILLIN

G

MODULE

NETWORK

MODULE

LIBPCAP

BPF

TCPDUMP

WINPCAP

PACKET FILTER

WINDUMP

DAG2, DAG3E SYSTEMS

DAGTOOLS

DAGDUMP

ETHERNET, FDDI, ATM, SONET

10/100 Mbps, DS3, OC3, OC12

OS: LINUX/UNIX OR WINDOWS

Billing data

CHARGING MODELS

2nd Moment Model

Effective Bandwidth

Token Bucket…

Models parameters definition

Link capacity, buffer size, window size…

Timestamp & byte count samples.

“Scalars” per unit of time, constants and billing period.

Billing parameters definition:

Charging equation, price units, billing period, peak hours/days, non peak hours/days, minimum rate, etc.

Bill report creation (File creation and storing)

C++

C++

Database Manager: Access or Oracle, etc.

Timestamp granularity definition.

ETHERNET

10/100 Mbps

ETHERNET

10/100 Mbps

COMPUTATIO

N

MODULE

BILLIN

G

MODULE

NETWORK

MODULE

LIBPCAP

BPF

TCPDUMP

WINPCAP

PACKET FILTER

WINDUMP

DAG2, DAG3E SYSTEMS

DAGTOOLS

DAGDUMP

ETHERNET, FDDI, ATM, SONET

10/100 Mbps, DS3, OC3, OC12

OS: LINUX/UNIX OR WINDOWS

Billing data

CHARGING MODELS

2nd Moment Model

Effective Bandwidth

Token Bucket…

Models parameters definition

Link capacity, buffer size, window size…

Timestamp & byte count samples.

“Scalars” per unit of time, constants and billing period.

Billing parameters definition:

Charging equation, price units, billing period, peak hours/days, non peak hours/days, minimum rate, etc.

Bill report creation (File creation and storing)

C++

C++

Database Manager: Access or Oracle, etc.

Timestamp granularity definition.

ETHERNET

10/100 Mbps

ETHERNET

10/100 Mbps

Effective Bandwidth Model

• Rp peak rate, utilization, b mean burst period, B buffer

size, buffer overflow probability.

• Key variable: This is the tradeoff between bandwidth and buffer. (users can purchase buffer capacity from ISP).

• a scalar.

• Provides incentive for appropriate buffer size selection

expostP *( )aCB

2(1 ) ( (1 ) ) 4 (1 )

2 (1 )p p pb R B b R B B b R

Cb

ln(1 )

Price vs. Buffer Size (Same Trace / Different Deltas)

Price vs. Buffer Size (Same Trace / Different Deltas)

0

5

10

15

20

25

30

0 10 20 30 40 50 60 70 80 90 100

Buffer Size(Mbits)

Pri

ce

Price vs. Buffer SizePrice vs. Buffer Size

10.8

11.0

11.2

11.4

11.6

11.8

12.0

12.2

12.4

12.6

0 20 40 60 80 100

Buffer Size(Mbits)

Pri

ce

Price vs. Buffer Size(Same Delta / Different Traces)

Price vs. Buffer Size (Same Delta / Different Traces)

0

10

20

30

40

50

60

70

80

90

0 10 20 30 40 50 60 70 80 90 100

Buffer Size (Mbits)

Pri

ce

43.9%

26.3%

6.8%

Price vs. Time

Price vs. Time

0

2

4

6

8

10

12

14

0 10 20 30 40 50 60 70 80 90 100

Time (sec)

Pri

ce

Moment-Based Model

• x is original trace. y is smoothed trace, obtained by taking a simple moving average (with moving average window size w).

utilization. B bandwidth of link. a, b, c, scalars.• Uses basic statistical measures. Easy to understand

for non-engineers.

Pex-post = a Ln (b + 1) ec r2+l2 - 1

B

xr

]E[ 2

B

yl

]E[ 2

Moment-Based Model:Window Size Sensitivity

Price vs. Window Size

0.009

0.0095

0.01

0.0105

0.011

0.0115

0.012

0 50 100

Window Size

Pri

ce Price

Moment-Based Model:Price Convergence

0

0.001

0.002

0.003

0.004

0.005

0.006

0.007

0.008

0.009

0 10 20 30 40 50 60 70 80 90 100Time (s)

Pric

e

Ex-Post Charging Diffusion

• user i, a light user of Internet resources, who currently subsidizes some of the heavier users, switches from an ex-ante charging model to an ex-post charging model to lower their costs.

• user j, a medium user of Internet resources, is now faced with larger costs because they are no longer subsidized by user i, chooses to manage their network better and adopt an ex-post charging model.

• user k, a heavy user of Internet resources, is now faced with larger costs because they are no longer subsidized by users i and j, so they switch to an ex-post charging model.

Quality of Service

Pexpost = a1*(C1 + 1B1) + a2*(C2 + 2B2)• Users get more specific about QoS needs, then ex-

post algorithms could be more specific (SLAs)• But, to have an environment where guaranteed

end-to-end QOS applications can be provided over the internet there must be overall adoption of better network management among all users.

• Ingress and Egress is where the bottlenecks are and this is where the charging algorithms will be working

Security (Benefits)

• Accounting gives network managers metrics to identify potential security problems

• Charging based on these metrics gives network managers an incentive to monitor their networks.

Research Dissemination

• Ex-Post Internet Charging, J. Bailey, J. Nagel, and S. Raghavan to appear in Internet Services, L. McKnight and J. Wroclawski, eds., MIT Press

• Ex-Post Internet Charging Models Using Effective Bandwidth, I. Gamvros, C. Hernandez, J. Bailey, S. Raghavan, working paper to be completed summer 2001.

• A Prototype Ex-Post Internet Charging System, J. Noda, S. Raghavan, and J. Bailey, working paper to be completed shortly.

• Second Moment Model for Ex-Post Internet Charging, C. Lapuerta, J. Bailey, and S. Raghavan, working paper to be completed summer 2001.

Conclusions

• Simple pricing scheme, implementable in practice, provides incentives to users to manage and monitor their networks. Result is a more stable/secure network where it is easy to provide QOS guarantees.

• Pushes responsibility to edges of networks..• Determination of scalars. Develop (ISP) cost

models to provide guidelines on how to set scalars.