56
1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical Engineering and Computer Science Department University of California, Berkeley Berkeley, CA 94720-1776

1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

  • View
    213

  • Download
    0

Embed Size (px)

Citation preview

Page 1: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

1

Berkeley-Helsinki Short Course

Lecture #8a: Auction Design and Implementations for

Network Resource Allocation

Weidong Cui and Randy H. KatzElectrical Engineering and Computer Science Department

University of California, BerkeleyBerkeley, CA 94720-1776

Page 2: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

2

Outline

• Motivation• Auction Design• Auction Implementations• Auction-based Applications for Resource

Allocation• Bandwidth Trading• Open Issues in Electronic Auctions for

Network Resource Allocation

Page 3: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

3

Outline

• Motivation• Auction Design• Auction Implementations• Auction-based Applications for Resource

Allocation• Bandwidth Trading• Open Issues in Electronic Auctions for

Network Resource Allocation

Page 4: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

4

Motivation

• Auctions are one of the oldest form of markets! (dating back to 500BC)

• A variety of commodities are sold using auctions today!– Spectrum (FCC)– Oil (OPEC)

• Automated auctions using software agents fit the requirements of network resource allocation very well.– Distributed selfish participants– Fine time/space granularity

Page 5: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

5

Outline

• Motivation• Auction Design• Auction Implementations• Auction-based Applications for Resource

Allocation• Bandwidth Trading• Open Issues in Electronic Auctions for

Network Resource Allocation

Page 6: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

6

Auctions - Definition

• A definition by McAfee and McMillan– An auction is a market institution with

an explicit set of rules determining resource allocation and prices on the basis of bids from the market participants.

Page 7: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

7

Auction Design

• Auctions are the main focus of economic mechanism design.

• Economic mechanism design– Design “rules of interaction” for

economic transactions that will yield some desired outcome.

• Tools– Microeconomics– Game Theory

Page 8: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

8

Main Components of Auctions• Bidding

– How to express bids efficiently?– How to do auction communication?

• Allocation– How to allocate resource once the bids are all in.

• Payment– How much does each winner of the auction pay?

• Strategy– Each (selfish) bidder is free to choose an arbitrary

bidding strategy once the auction’s protocol, allocation and payment rules are fixed.

– For some mechanisms, truth-telling is the dominant strategy.

Page 9: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

9

Bidding

• Bids– For combinatorial auctions, there are an exponential

number of bid combinations.– Bids may be coded by the auction communication.

• Price-quotes– Auctioneer send intermediate auction results to

bidders.– The format of price-quotes is dependent on auction

type and bidding state.

• Other interactions– Bid withdrawal/Admittance/Rejection– Transaction Notification

Page 10: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

10

Bidding

• Efficiency is the key problem for bidding!

• Price-quantity graphs– Bidders can express continuous preferences.– It’s useful for divisible resources, e.g.,

bandwidth.

• OR bids or XOR bids– For combinatorial auctions

• Code Bids• Reduce bidding rounds

Page 11: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

11

Allocation and Payment• Economic Efficiency

– Social Welfare Maximization– Seller Profit Maximization

• Computational Efficiency– Compute allocation results and payment in

polynomial time– Generally, it’s NP-hard for combinatorial auctions.

• Incentive Compatible– Bidders optimize their expected utilities by

bidding their true valuations for the good.

• Avoid Collusion

Page 12: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

12

Allocation and Payment

• The trade-off between economic efficiency and computational efficiency is constrained by the underlying network technology.– How much measurement (from usage to capacity

pricing)– The granularity of differently priced service offerings

(e.g., number of traffic classes)– The level of resource aggregation – both in time and in

space – at which pricing is done (per packet/cell or per connection, at the edge of the network or at each hop)

– The information requirement (how much a priori knowledge of user behavior and preferences is required/assumed by the network in computing prices)

Aurel A. Lazar and Nemo Semret“Design and Analysis of the Progressive Second Price Auction for Network Bandwidth Sharing”.

Page 13: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

13

Bidder’s Strategy

• Truth-telling (Risk Neutral)– In incentive-compatible auctions, truth-

telling is the dominant strategy.

• Collusion– Some subset of bidders coordinate their

bids to gain more value.

• Risk Aversion– Bidders are likely to raise their bids so that

they are more likely to win.

• Bid Shading

Page 14: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

14

Auctions - Taxonomy

• Criteria– single/double-sided– outcry/sealed-bid– ascending/descending bids– first price/second price– discriminatory/uniform price– single-object/combinatorial– single-unit/multi-unit– indivisible/divisible resources– close at once/continuous-bid– independent-private-value/common-value

Page 15: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

15

Main Auction Types

single double

SB outcry

ascending descending SB outcry

FPSBVickrey

DutchEnglishClearingHouse

CDA

P.R. Wurman, M.P. Wellman, and W.E. Walsh,“The Michigan Internet AuctionBot: A Configurable Auction Service for Human and Software Agents”, 1998.

Page 16: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

16

Outline

• Motivation• Auction Design• Auction Implementations• Auction-based Applications for Resource

Allocation• Bandwidth Trading• Open Issues in Electronic Auctions for

Network Resource Allocation

Page 17: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

17

Auction Implementations

• The Michigan Internet AuctionBot– Wurman, Wellman, and Walsh, University of

Michigan.– A configurable auction service for human and

software agents

• eMediator– Sandholm, Washington University.– A next generation electronic commerce server

• A Secure Electronic Auction Protocol– Srividhya Subramanian, The Ohio State

University

Page 18: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

18

The Michigan Internet AuctionBot

HTTPServer

TCP Server

CGI

CGI

Database

Scheduler

Auctioneer

E-mailServer

Computer

Web Interface

Agent Interface

The AuctionBot Architecture

P.R. Wurman, M.P. Wellman, and W.E. Walsh,“The Michigan Internet AuctionBot: A Configurable Auction Service for Human and Software Agents”, 1998.

Page 19: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

19

The Michigan Internet AuctionBot

• The AuctionBot supports a wide range of auction types by decomposing the auction design space into a set of orthogonal parameters.– Bidding Restrictions– Auction Events– Information Revelation– Allocation Policies

Page 20: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

20

The Michigan Internet AuctionBot

• Bidding Restrictions– Participation

• Single seller, multiple buyers• Multiple sellers, single buyer• Multiple sellers, multiple buyers

– Discrete Goods• Reject bids for non-integer quantities

– Bid Rules• Ascending• Descending• No withdrawal

Page 21: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

21

The Michigan Internet AuctionBot

• Auction Events– Clearing Schedule/Quote Schedule

• At scheduled times• At random times• By bidder activity• By bidder inactivity

– Closing Conditions• At a scheduled time• At a random time• After a period of inactivity• When designated bids are matched

Page 22: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

22

The Michigan Internet AuctionBot

• Information Revelation– Price Quotes

• Bid quote: the highest price to sell• Ask quote: the lowest price to buy

– Transaction History• Whether publicize selected information about

past transactions or not

– Schedule Information• Whether reveal the timing of upcoming clear and

quote events or not.

Page 23: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

23

The Michigan Internet AuctionBot

• Allocation Policies– Currently support three policies.– All are uniform-price and for discrete goods.– Mth-price policy/(M+1)st-price policy

• M is the number of units offered for sale.• Set the price at the Mth/(M+1)st highest among

all bids.• When M=1, Mth-price policy is a first-price auction

and (M+1)st-price policy is a second-price auction.

– Chronological match policy• Sort of Continuous Double Auction

Page 24: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

24

The Michigan Internet AuctionBot

• Real Time Issues– The bidder interface is decoupled from the

auction processing completely.– There is asynchrony between the scheduler

and the interface.– Solution: keep track of the state of a bid

• Unprocessed, valid, rejected, expired, partially transacted, transacted, replaced, withdrawn-requested, withdrawn.

• Denial-of-Service Attack– What if some software agents submit bids and

information requests at high frequency?

Page 25: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

25

eMediator

• eAuctionHouse• Leveled Commitment Contract

Optimizer• eExchangeHouse

Tuomas Sandholm, “eMediator: A Next Generation Electronic Commerce Server”, 2000.

Page 26: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

26

eMediator

• eAuctionHouse– A variety of generalized combinatorial

auctions– Bidding via graphically drawn price-quantity

graphs– Mobile agents

Page 27: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

27

eMediator

RegularPrice Bids

Price-QuantityGraph Bids

OR-XOR-Bids

OR-Bids

Single auction,multiple items,multiple units of each

Single auction,1 unit of 1 item

Single auction,multiple units of 1 item

Single auction,multiple items, 1 unit of each

Double auction,multiple items,multiple units of each

Double auction,1 unit of 1 item

Double auction,multiple units of 1 item

Double auction,multiple items, 1 unit of each

First-Price

MiddlePrice

(50:50)

2nd-Price(Vickrey)

Multi-unitVickrey

Groves

Bid Type Auction Setting Pricing Scheme

Page 28: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

28

eMediator• Mobile agents as auction participants

– Information agent• Informs users auction progress

– Incrementor agent (for English auction)• Bids a small increment more than the current highest

price• Stops if the user’s reservation price is reached

– N-agent• Underbids when the number of bidders is N.• Bids the user’s valuation times (N-1)/N.

– Control agent• Submit very low noncompetitive bids• Increases the number of bidders• Misleads N-agents

– Discover agent• Computes the expected gain from bidding a small

increment more than the current highest price according to the agent’s current distribution of her valuation

Page 29: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

29

eMediator

• Leveled Commitment Contract Optimizer– Full commitment contract are unable to take

advantage of the possibilities that such future events provide.

– Contingency Contracts• The contract obligations are made contingent on

future events.

– Leveled Commitment Contracts• The level of commitment by decommitment

penalties are specified in the contract.

Page 30: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

30

eMediator• eExchangeHouse

– A safe exchange planner.– Make sure that the seller gets paid and the

buyer gets the good.– Approach

• Divide the exchange into chunks where each party delivers a small amount at a time,

• and the exchange proceeds with such alternation.

– A sequence is called safe if each party is motivated to follow the exchange at every step in anticipation of the profit from the rest of the exchange instead of vanishing with what the other party has delivered so far.

Page 31: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

31

A Secure Electronic Auction Protocol

• Security Issues in Electronic Auctions– Anonymity of bidders– Security from passive attacks, active attacks,

message corruption, and loss of messages– Bidder’s privacy– Atomicity of transaction

• By using a logic developed based on the semantics of BAN-style logic, Subramanian proves that the proposed secure electronic auction protocol ensures all the properties.

Srividhya Subramanian,“Design and Verification of a Secure Electronic Auction Protocol”, 1996.

Page 32: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

32

Outline

• Motivation• Auction Design• Auction Implementations• Auction-based Applications for Resource

Allocation• Bandwidth Trading• Open Issues in Electronic Auctions for

Network Resource Allocation

Page 33: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

33

FCC Spectrum Auctions

• The US government sold spectrum right using an innovative auction design, the simultaneous ascending auction.– Bidders bid on numerous communication licenses

simultaneously, with bidding remaining open on all licenses until no bidder is willing to bid higher on any licenses.

• Collusive bidding in FCC spectrum auctions– Bidders send messages to their rivals, telling them on

which licenses to bid and with to avoid.– These strategies can help bidders coordinate a division

of the licenses, and enforce the proposed division by directed punishments.

John McMillan, “Selling Spectrum Rights”, 1994.Peter Cramton and Jesse A. Schwartz,“Collusive Bidding: Lessons from the FCC Spectrum Auctions”, 1999.

Page 34: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

34

Bandwidth Allocation Over Paths

• A set of simultaneous multi-unit descending auctions, one per link of the network.

• To win bandwidth over a certain path, it suffices to simultaneously bid for the quantity desired at all relevant auctions.

• Prices at the various links drop at different rates, following specified rules so that prices reflect the demand exhibited for each link.

• A Vickrey-type pricing rule is used to address the issue of incentive compatible pricing.

• Problems:– How to break tie when two bids arrive almost

simultaneously?– How to control the period of closing auctions?

Costas Courcoubetis, Manos P. Dramitinos, and George D. stamoulis,“An Auction Mechanism for Bandwidth Allocation Over Path”, 2000.

Page 35: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

35

“Smart Market”

• The price to send a packet would vary minute-by-minute to reflect the current degree of network congestion.

• No charge for sending packets when the networks is not congested.

• Per-packet charge when the network is congested.• Users only pay the market-clearing price, which is

always lower than the bids of all admitted packets.• This is actually a “second-price” auction.• Problem

– Not Scalable!

Jeffrey K. MacKie-Mason and Hal R. Varian, “Pricing the Internet”, 1994.

Page 36: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

36

Congestion-based Pricing• Critical resource reaches congestion levels,

modify prices to drive utilization back to “acceptable” levels.

• Berkeley Computer Telephony Service Testbed– Gateways as bottlenecks (limited PSTN access lines)– Use congestion pricing (CP) to entice users to

• Talk shorter• Talk later• Accept lower quality

– Auction type???

• Upenn Modem Pool– Mth-price auction

Jimmy Shih and Randy H. Katz, “???”, 2001.Frank J. Klausz, David C. Croson, and Rachel T.A. Croson,“An Experimental Auction to Allocate Congested IT Resources:The Case of the Universtiy of Pennsylvania Modem Pool”, 1998.

Page 37: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

37

Progressive Second Price Auction for Network Bandwidth

Sharing• PSP is based on an exclusion-compensation

principle– Each bidder pays for his allocation so as to exactly cover

the “social opportunity cost” which is given by the declared willingness to pay (bids) of the bidders who are excluded by his presence.

• PSP is incentive compatible and stable under elastic demand.

• PSP is economically efficient in that the equilibrium allocation maximizes total user value.

• Problems– It can not be applied directly to path allocation in

networks!

Aurel A. Lazar and Nemo Semret,“Design and Analysis of the Progressive Second Price Auction for Network Bandwidth Sharing”, 1999.

Page 38: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

38

Charge-Sensitive TCP and Rate Control

• How to achieve the system optimal rates in a distributed environment, which maximize the total user utility, using only the information available at the end hosts?– Decompose the system problem into two

subproblems: network and user problems,– Introduce an incentive-compatible pricing scheme,

while maintaining proportional fairness.

• It’s demonstrated that, when users update their parameters by solving their own optimization problem, at an equilibrium the system optimum is achieved.

Richard J. La and Venkat Anantharam,“Charge-Sensitive TCP and Rate Control in the Internet”, 2000.

Page 39: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

39

Market-based Routing in Telecommunication Networks• Economic agents

– Call agent– Path agent– Link agent

M.A. Gibney and N.R. Jennings,“Dynamic Resource Allocation by Market-based Routing in Telecommunication Networks”, 1998.

• Market institutions– Path market– Link market

CallAgents

LinkAgents

PathAgents

PathMaket

LinkMaket

Buyer SellerSeller Buyer

• This architecture shows a possible solution!

Page 40: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

40

Outline

• Motivation• Auction Design• Auction Implementations• Auction-based Applications for Resource

Allocation• Bandwidth Trading• Open Issues in Electronic Auctions for

Network Resource Allocation

Page 41: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

41

Bandwidth Trading• History

– Enron completed its first bandwidth trade with Global Crossing in December 1999.

– About 1000 bandwidth trades took place in 2000, about one-third of which involved Enron as a counterpart.

• Why use bandwidth trading?– The current negotiate market for

exchanging broadband capacity is too cumbersome and costly!

– Other commodity (e.g., energy) trading is successful!

Page 42: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

42

Bandwidth Trading

• Prices going down, competition going up

Page 43: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

43

Bandwidth Trading

• Players– Market Makers/Traders

• Enron• Williams Communications• Dynergy• LightTrade

– Bandwidth Exchanges• Arbinet• Band-X• Ratexchange• Commerex

Page 44: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

44

Bandwidth Trading

• Requirements for making bandwidth a commodity– Bandwidth liquidity– Price transparency– Standard contracts– Liquidated damages– Common quality benchmarks

• A Bandwidth Trading Organization (BTO) trading agreement is expected by the end of this year!

Page 45: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

45

Pooling Points

Source: Enron Broadband Services

• Enron Pooling Point Network (PPN)

Page 46: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

46

Pooling Points• Primary function is

– to provide dynamic, real-time provisioning and delivery of bandwidth between buyers and sellers, and

– to provide the QoS measurements necessary to create truly fungible units of capacity.

• One version of a pooling point is a high-capacity switch connected to a network element of each participating capacity buyer and seller.– Demarcation point– I/O ports– Switch matrix

Page 47: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

47

An Example ofan Interconnection Option

Source: Enron Broadband Services

Page 48: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

48

Pooling Points

• Enron– 2000: 21 pooling points– 2001: 35 pooling points (expected)– Enron claims the ability of change circuits

every 5 seconds.– Based on Lucent network management

technology

• LightTrade– 2000: 8 pooling points– 2001: 15 pooling points (expected)– Neutral pooling points

Page 49: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

49

Other Interconnection Method

• Williams Communications is already connected to most carriers through its own links.

• It doesn’t have to “actively” seek a cross-connect by hooking up to pooling points.

Page 50: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

50

Where do exchanges happen?

• Current exchanges can only function at Layers 1 and 2.

• Providers are loath to automate the BGP peering sessions needed to establish Layer 3 IP connectivity.– Routing tables are not stable.– The way the BGP peering sessions work

across networks is volatile.– The lack of end-to-end QoS and the inter-

domain issues of MPLS.

Page 51: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

51

What do we not know about bandwidth trading?

• Can the bandwidth trading be automated?

• What mechanism is used to clear the market?

• How’s the negotiation process performed?

Page 52: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

52

Outline

• Motivation• Auction Design• Auction Implementations• Auction-based Applications for Resource

Allocation• Bandwidth Trading• Open Issues in Electronic Auctions for

Network Resource Allocation

Page 53: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

53

Open Issues in Electronic Auctions

for Resource Allocation• How can we do bandwidth trading at

Layer 3 at a fine grain time scale in wide-area networks?

• In general, Combinatorial auctions are NP-hard. Is it possible to apply combinatorial auctions to path allocation in real time?

• Collusion detection and avoidance is very important for real-world bandwidth trading. How can we solve this problem?

Page 54: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

54

Open Issues in Electronic Auctions

for Resource Allocation• Is hierarchical auction a good way to allocate

resource in wide-area networks?– A few small auctioneer distributes in the wide-area

network.– Local users submit bids (resource requests) to the

local auctioneer.– The local auctioneer aggregates users’ requests and

submit bids to a global auctioneer.– The global auctioneer allocates resources (or

exchange resources) and inform local auctioneers.– How local auctioneers function is a challenging

problem.

Page 55: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

55

Open Issues in Electronic Auctions

for Resource Allocation• User’s utility is mentioned in every paper.

What is user’s utility in real network environment?

• What is the requirement for end-to-end QoS if we want to implement real end-to-end bandwidth allocation?

• Can electronic auctions for resource allocation take advantage of achievements of overlay networks?– Decentralized auctioneers construct an

overlay network!

Page 56: 1 Berkeley-Helsinki Short Course Lecture #8a: Auction Design and Implementations for Network Resource Allocation Weidong Cui and Randy H. Katz Electrical

56

Open Issues in Electronic Auctions

for Network Resource Allocation• For the auction communication in wide-

area networks, can we learn some lessons from wide-area signaling protocols?

• Is there any other mechanism better than auctions for network resource allocation?