EECS228a Lecture 2 Research Topics Jean Walrand - - PowerPoint PPT Presentation

EECS228a – Lecture 2 Research Topics

Jean Walrand www.eecs.berkeley.edu/~wlr

Walrand EECS 228a 52

Outline

Economics of Networks Routing Congestion Control Traffic Models

Walrand EECS 228a 53

Economics of Networks

Outline

Hangover Pricing of Services Competition of Users Competition of Providers Suggested Readings:

n http://www.bgsu.edu/departments/tcom/annota.htm n http://info.isoc.org/internet-history/ n http://www.spp.umich.edu/ipps/papers/info-

nets/Economic_FAQs/FAQs/FAQs.html

Walrand EECS 228a 54

Economics of Networks

Hangover

Bubble: Wired

Walrand EECS 228a 55

Economics of Networks

Hangover

Bubble: Wireless

Walrand EECS 228a 56

Economics of Networks

Hangover

Over-Investment

n Based on unrealistic growth forecast n Overcapacity: Fiber

5x100 in three years

n Too many companies competing for same market

Debt

n Wireless: Expensive spectrum licenses n Fibers n IT in companies: PCs, Servers, Networks

Walrand EECS 228a 57

Economics

Key Ideas

Value of services to users: externality, QoS, CoS Market segmentation Flat rate pricing; congestion pricing; Paris metro pricing; time-of-day pricing Incentive compatibility Inter-ISP settlements; Peering agreements Internet as a public good

Walrand EECS 228a 58

Economics

Value of Services

Externality: Kazaa Value per bit: email vs. fax vs. picture Value of bit rate: video stream vs. radio Value of low latency: video stream vs. video conference Value of low response time: browsing with DSL vs. browsing with 56k QoS affects value and usage Value of QoS depends on application and user

Walrand EECS 228a 59

Economics

Market Segmentation

Businesses vs. Residential Customers Network Application Providers vs. public Web Sites Principle: Charge more users with higher utility

Walrand EECS 228a 60

Economics

Differentiated Pricing

Examples:

n First Class & Economy in plane: More space but much more

expensive

n Paris Metro: More expensive

Fewer Users Better Service (e.g., Stanford vs. Berkeley?)

Suggests Class of Service:

n Better service by mechanism: e.g., priority n Better service by fewer users: e.g., expensive network;

congestion pricing (e.g., packet marking); time-of-day

Alternative: QoS: You know what you pay for

n Service Level Agreement (implementation?) n QoS of accepted calls: end-to-end test

Walrand EECS 228a 61

Economics

Incentive Compatibility

How to discover the user’s willingness to pay? Examples:

n California Electricity: Providers offer bids and CA

buys cheaper first prices escalade

n Highest bidder auction: Spectrum auctions n Highest gets but two highest pay n Second highest price: Incentive compatible

Walrand EECS 228a 62

Economics

Competition

Basic supply and demand:

n More capacity than traffic

prices drop and providers go bankrupt

Internet traffic doubles every year instead of every 100 days …. Quality service is still rare and valuable:

n Businesses use video conference over ISDN n Expensive commutes and business travel n Users pay a lot for CATV and pay-per-view n T1 service expensive: demand exists

Walrand EECS 228a 63

Economics

Game Theory

Framework to analyze result of interaction of self-interested agents Suggests strategies for

n Pricing services n Peering agreements n Routing n QoS definitions n Evolution of industry (e.g., consolidation vs.

specialization)

Two parts: Games & Mechanism Design

Walrand EECS 228a 64

Routing

Outline

Motivation Granularity Types Issues

Walrand EECS 228a 65

Routing

Motivation

Reduce delays: Avoid OAK NY SF Improve reliability: Protection Sensor networks: Many open questions Ad Hoc networks: More robust, provide QoS IP/Optical: Improve coordination

Walrand EECS 228a 66

Routing

Granularity

Light Path: WDM Cross-Connect: SONET Circuit: Telephone Label Switched Path: MPLS; ATM Connection Packet

Walrand EECS 228a 67

Routing

Granularity (cont)

Benefit of LSP vs SONET is not obvious:

n Consider traffic from SF to NY; If that

traffic is essentially constant, then SONET is good enough. If not, LSP/SONET is preferable.

n If traffic is self-similar, then fluctuations

persist at high rate

Walrand EECS 228a 68

Routing

Types

On-line vs. Off-line Centralized vs. Distributed Link State; Distance Vector; Path Vector Source-based vs. Destination-based QoS routing Ad Hoc; Location-Based Ant-routing (reinforcement) Unicast vs. multicast Protection routing Peer-to-peer vs. overlay

Walrand EECS 228a 69

Routing

Issues

Benefits Implementability:

n Scalability: communications required;

complexity; convergence time

n Robustness: sensitivity to errors

Walrand EECS 228a 70

Congestion Control

Outline

Motivation Examples Issues

Walrand EECS 228a 71

Congestion Control

Motivation

At user level: Issues with QoS At network level: Losses, inefficiency, unfairness At switch level: Scalability problems

Walrand EECS 228a 72

Congestion Control

Examples

TCP Congestion in routers Call Admission Control

Walrand EECS 228a 73

Congestion Control

Issues

Fairness vs. Optimality Simplicity Robustness

Walrand EECS 228a 74

Traffic Models

Outline

Why bother? Transactions Packet flows

Walrand EECS 228a 75

Traffic Models

Why Bother?

Network should be robust; not based on detailed traffic assumptions Traffic characteristics impact

n Effectiveness of multiplexing n Buffer sizes required n Time scale of bandwidth allocations

Walrand EECS 228a 76

Traffic Models

Transactions

File transfers:

n File sizes: Heavy tailed n Timing of requests: Poisson n Geography:

w Kazaa – poor locality w Akamai – improved locality

Other applications:

n video conferences n VoIP

Walrand EECS 228a 77

Traffic Models

Packet Flows

Self-Similarity:

n Heavy Tail + TCP

Self Similar Flows

n Heavy Tail Files + Structure of Web Sites

Self Similarity

Relevance:

n Not obvious – a matter of time scale