online procurement auctions for resource pooling in client-assisted cloud storage systems jian zhao,...

27
Online Procurement Auctions for Resource Pooling in Client-Assisted Cloud Storage Systems Jian Zhao, Xiaowen Chu , Hai Liu, Yiu-Wing Leung Department of Computer Science Hong Kong Baptist University Zongpeng Li Department of Computer Science University of Calgary IEEE INFOCOM 2015, HONG KONG 1

Upload: sheryl-flynn

Post on 24-Dec-2015

218 views

Category:

Documents


1 download

TRANSCRIPT

Page 1: Online Procurement Auctions for Resource Pooling in Client-Assisted Cloud Storage Systems Jian Zhao, Xiaowen Chu, Hai Liu, Yiu-Wing Leung Department of

IEEE

INFO

COM

201

5, H

ON

G K

ON

G

Online Procurement Auctions for Resource Pooling

in Client-Assisted Cloud Storage Systems

Jian Zhao, Xiaowen Chu, Hai Liu, Yiu-Wing LeungDepartment of Computer ScienceHong Kong Baptist University

Zongpeng LiDepartment of Computer ScienceUniversity of Calgary 1

Page 2: Online Procurement Auctions for Resource Pooling in Client-Assisted Cloud Storage Systems Jian Zhao, Xiaowen Chu, Hai Liu, Yiu-Wing Leung Department of

IEEE

INFO

COM

201

5, H

ON

G K

ON

G

Outline

Motivation: why client-assisted?

Background: auctions

Model formulation

Some technical details (without equations)

Conclusions 2

Page 3: Online Procurement Auctions for Resource Pooling in Client-Assisted Cloud Storage Systems Jian Zhao, Xiaowen Chu, Hai Liu, Yiu-Wing Leung Department of

IEEE

INFO

COM

201

5, H

ON

G K

ON

G

Cloud Storage Services

3

Page 4: Online Procurement Auctions for Resource Pooling in Client-Assisted Cloud Storage Systems Jian Zhao, Xiaowen Chu, Hai Liu, Yiu-Wing Leung Department of

IEEE

INFO

COM

201

5, H

ON

G K

ON

G

4Durability: Our data will be there without error forever.

Most of the time, CSPs are

Availability: Data can be accessed anywhere, anytime, from any device.

Page 5: Online Procurement Auctions for Resource Pooling in Client-Assisted Cloud Storage Systems Jian Zhao, Xiaowen Chu, Hai Liu, Yiu-Wing Leung Department of

IEEE

INFO

COM

201

5, H

ON

G K

ON

G

5

But occasionally, they become

Cloud Outages!

Page 6: Online Procurement Auctions for Resource Pooling in Client-Assisted Cloud Storage Systems Jian Zhao, Xiaowen Chu, Hai Liu, Yiu-Wing Leung Department of

IEEE

INFO

COM

201

5, H

ON

G K

ON

G

Cloud OutagesCloud services could become unavailable because of

6

Page 7: Online Procurement Auctions for Resource Pooling in Client-Assisted Cloud Storage Systems Jian Zhao, Xiaowen Chu, Hai Liu, Yiu-Wing Leung Department of

IEEE

INFO

COM

201

5, H

ON

G K

ON

G

Evidence of Cloud Outages

• “The worst cloud outagesof 201X “ by J. R. Raphael

Inforworld.com has been

tracing cloud outages since

2011

• Amazon• Google• Microsoft• Apple• Dropbox• Facebook• Adobe• …

Big names on the list

7

Page 8: Online Procurement Auctions for Resource Pooling in Client-Assisted Cloud Storage Systems Jian Zhao, Xiaowen Chu, Hai Liu, Yiu-Wing Leung Department of

IEEE

INFO

COM

201

5, H

ON

G K

ON

G

What can we do?Plan A: Cloud Federation

8

Page 9: Online Procurement Auctions for Resource Pooling in Client-Assisted Cloud Storage Systems Jian Zhao, Xiaowen Chu, Hai Liu, Yiu-Wing Leung Department of

IEEE

INFO

COM

201

5, H

ON

G K

ON

G

Alternatively

9

Data Center A

Data Center B

Data Center C

storage pooling

storage retail

storage pooling

storage retail

high speed network

Plan B: Client-Assisted Cloud Storage

Page 10: Online Procurement Auctions for Resource Pooling in Client-Assisted Cloud Storage Systems Jian Zhao, Xiaowen Chu, Hai Liu, Yiu-Wing Leung Department of

IEEE

INFO

COM

201

5, H

ON

G K

ON

G

Client-Assisted Examples in Academia

• Leverage peer bandwidth to mitigate server bandwidth cost

• Improve availability and downloading performance

FS2You

• Data is kept at a central cloud and replicated among distributed peersAmazingStore

• Peer-assisted architecture with a focus on data consistencyTriton

10

Page 11: Online Procurement Auctions for Resource Pooling in Client-Assisted Cloud Storage Systems Jian Zhao, Xiaowen Chu, Hai Liu, Yiu-Wing Leung Department of

IEEE

INFO

COM

201

5, H

ON

G K

ON

G

A Counter Example:

Wuala abandoned the “client-assisted” design in 2012

Hybrid architecture is very complicated

Bandwidth cost is dropping

Contribution from peers are marginal

More focus on business customers

Wuala was designed to be “client-assisted”.Servers in datacenters

Client users• casual peers without contribution• storage peers trading local storage for increased

storage space

11

Page 12: Online Procurement Auctions for Resource Pooling in Client-Assisted Cloud Storage Systems Jian Zhao, Xiaowen Chu, Hai Liu, Yiu-Wing Leung Department of

IEEE

INFO

COM

201

5, H

ON

G K

ON

G

Another Story: Symform

Observation•“cloud providers use extremely inefficient centralized infrastructure to store stuff.”

•“most users had tons of excess local storage just going to waste”

Goal•“Creating the World’s Largest Datacenter”

•“Without Building a Datacenter”

Method•“Users that contribute get 1 GB free for every 2 GB contributed”

12

Acquired by Quantum in 2014

Page 13: Online Procurement Auctions for Resource Pooling in Client-Assisted Cloud Storage Systems Jian Zhao, Xiaowen Chu, Hai Liu, Yiu-Wing Leung Department of

IEEE

INFO

COM

201

5, H

ON

G K

ON

G

Incentive makes a difference

Public BitTorrent

• Tit-for-Tat only• Free-rider problem• Stop seeding after

downloading• Limit the uploading

bandwidth

Private BitTorrent (or Darknet)

• Sharing Ratio Enforcement• Users fight to contribute

as much as they can to survive

• Everyone in the community can get high downloading performance

• People are willing to pay to get into the community

13

Total upload = Total download

Page 14: Online Procurement Auctions for Resource Pooling in Client-Assisted Cloud Storage Systems Jian Zhao, Xiaowen Chu, Hai Liu, Yiu-Wing Leung Department of

IEEE

INFO

COM

201

5, H

ON

G K

ON

G

Auctions are economical approaches for allocating resources or trading commodities

Participants: auctioneer + bidders

Auction mechanism: a set of institutions for the buying/selling of goods or services, such as allocation rules and pricing rules

Auctions

14

Page 15: Online Procurement Auctions for Resource Pooling in Client-Assisted Cloud Storage Systems Jian Zhao, Xiaowen Chu, Hai Liu, Yiu-Wing Leung Department of

IEEE

INFO

COM

201

5, H

ON

G K

ON

G

Auctions as Incentive

No need to predict users’ demand: users reveal their true information through bids

Balance among supply and demand

Allocate available resources efficiently

Obtain higher revenue 15

Page 16: Online Procurement Auctions for Resource Pooling in Client-Assisted Cloud Storage Systems Jian Zhao, Xiaowen Chu, Hai Liu, Yiu-Wing Leung Department of

IEEE

INFO

COM

201

5, H

ON

G K

ON

G

Research Problem

How to design the auction mechanism for resource pooling from clients?

• For clients: get reasonable monetary return by selling resources (storage space & bandwidth)

• For CSPs: save cost (and overcome cloud outages)• Assumption: the cost of providing service by

datacenter herself is expensive than procuring resources from clients 16

Page 17: Online Procurement Auctions for Resource Pooling in Client-Assisted Cloud Storage Systems Jian Zhao, Xiaowen Chu, Hai Liu, Yiu-Wing Leung Department of

IEEE

INFO

COM

201

5, H

ON

G K

ON

G

Our Approach:Online Procurement Auctions

Online auctions

• Different bidders arrive at different times

• Auctioneer makes decision about each bid as it is received• Different from the

traditional case that the auctioneer receives all the bids before determining the allocation

• In line with asynchronous arrivals of user bids and requests

Procurement auctions

• Ordinary auctions (or forward auctions): one seller, multiple buyers• Buyers compete

• Procurement auctions: one buyer, multiple sellers• Sellers compete• reverse auctions

17

Page 18: Online Procurement Auctions for Resource Pooling in Client-Assisted Cloud Storage Systems Jian Zhao, Xiaowen Chu, Hai Liu, Yiu-Wing Leung Department of

IEEE

INFO

COM

201

5, H

ON

G K

ON

G

Our Auction Model (I)

Storage & bandwidth are unified as conceptual “resource”• Needs further study

Clients (i.e., bidders) want to sell resources through bids• Bid: (starting time, ending time, amount, unit price)• Valuation: the “true” value of the resource, private to the

bidder• Utility = the received payment - the cost of offering the sold

resource• Target: maximize utility

18

Page 19: Online Procurement Auctions for Resource Pooling in Client-Assisted Cloud Storage Systems Jian Zhao, Xiaowen Chu, Hai Liu, Yiu-Wing Leung Department of

IEEE

INFO

COM

201

5, H

ON

G K

ON

G

Our Auction Model (II)

CSP (i.e., auctioneer ) wants to buy resources from some clients• Set a target S for time period [0, T]• For each incoming bid, determine

• how much to procure (allocation rule)• how much to pay (payment rule)

19

Page 20: Online Procurement Auctions for Resource Pooling in Client-Assisted Cloud Storage Systems Jian Zhao, Xiaowen Chu, Hai Liu, Yiu-Wing Leung Department of

IEEE

INFO

COM

201

5, H

ON

G K

ON

G

Design Objectives

• If ( truthful bidding always maximizes utility )

• Then ( rational bidders will report their true valuations )

Truthful

• The CSP’s total cost is bounded by γ times of the total cost of offline optimal auction

Competitive20

Page 21: Online Procurement Auctions for Resource Pooling in Client-Assisted Cloud Storage Systems Jian Zhao, Xiaowen Chu, Hai Liu, Yiu-Wing Leung Department of

IEEE

INFO

COM

201

5, H

ON

G K

ON

G

Our Results

Truthfulness• We derive price-based allocation rule and

payment rule that result in truthful auction mechanism.

Competitiveness• We find a solution to guarantee a target

competitive ratio against offline optimal auctions.

21

Page 22: Online Procurement Auctions for Resource Pooling in Client-Assisted Cloud Storage Systems Jian Zhao, Xiaowen Chu, Hai Liu, Yiu-Wing Leung Department of

IEEE

INFO

COM

201

5, H

ON

G K

ON

G

Our Methodology

Start from Myerson’s Principles of Truthfulness (1981)•Gives two conditions to satisfy the truthfulness property

•For one single indivisible good

Extend to the case of online procurement auction for divisible goods•We first study the conditions for truthfulness of an online auction

•We then derive allocation and payment equations such that the auction mechanism is truthful.

Design marginal pricing function for CSP to achieve a desired competitive ratio

22

Page 23: Online Procurement Auctions for Resource Pooling in Client-Assisted Cloud Storage Systems Jian Zhao, Xiaowen Chu, Hai Liu, Yiu-Wing Leung Department of

IEEE

INFO

COM

201

5, H

ON

G K

ON

G

Conclusions

We consider “client-assisted” a promising approach to addressing cloud outages.

We argue that “auctions” can be a good incentive mechanism.

We propose an online procurement auction mechanism, and prove its truthfulness and competitiveness. 23

Page 24: Online Procurement Auctions for Resource Pooling in Client-Assisted Cloud Storage Systems Jian Zhao, Xiaowen Chu, Hai Liu, Yiu-Wing Leung Department of

IEEE

INFO

COM

201

5, H

ON

G K

ON

G

24

Aand

Page 25: Online Procurement Auctions for Resource Pooling in Client-Assisted Cloud Storage Systems Jian Zhao, Xiaowen Chu, Hai Liu, Yiu-Wing Leung Department of

IEEE

INFO

COM

201

5, H

ON

G K

ON

G

Myerson’s Principle of Truthfulness

Consider the auction of a single indivisible good

The auction mechanism is truthful if and only if• The probability of a bidder winning an auction

is monotonically non-decreasing in its bid• The payment charged to a bidder is

independent of its bid25

Page 26: Online Procurement Auctions for Resource Pooling in Client-Assisted Cloud Storage Systems Jian Zhao, Xiaowen Chu, Hai Liu, Yiu-Wing Leung Department of

IEEE

INFO

COM

201

5, H

ON

G K

ON

G

Extension to Online Procurement Auctions

We first define “allocation monotonicity” for online procurement auctions.• Better bids get more allocation

If allocation rule is monotone, we can find a payment rule that results in truthful online procurement auction.

We design a monotone allocation rule and the corresponding payment rule.

26

Page 27: Online Procurement Auctions for Resource Pooling in Client-Assisted Cloud Storage Systems Jian Zhao, Xiaowen Chu, Hai Liu, Yiu-Wing Leung Department of

IEEE

INFO

COM

201

5, H

ON

G K

ON

G

Competitive Analysis

Our auction mechanism assumes a non-increasing marginal pricing function for the CSP to procure resources from clients.• The marginal pricing function is a variable

How to minimize the total cost by adjusting the marginal pricing function?

• An online algorithm design problem

We find a solution of setting the marginal pricing function to achieve a target competitive ratio. 27