A Case Study on How Economic Frameworks

Can Bail Out Systems Research

Emin Gün SirerRyan Peterson, Bernard Wong

Department of Computer Science, Cornell UniversityUnited Networks, LLC

September 3, 2009

Problem Domain

What is the most efficient way to

disseminate a large set of files to

a large set of clients?

Client-Serverserver

clients

InefficientInefficientHigh cost of High cost of ownershipownership

Peer-to-Peerpeer

Limited informationLimited informationNo control or performance No control or performance

guaranteesguarantees

Traditional Systems Research

•Complex problems lead to complex protocols with many parameters

• Parameters picked based on a combination of gut feeling and experimentation

• Systems tailored to specific workloads

• Only intellectual honesty keeps researchers from perfect performance on benchmarks!

•Novelty by aggregation• Pick k out of n: CDNs, network coordinates,

geolocation, network tomography, …

BitTorrent

•Robust, resilient protocol• A combination of randomization and heuristics

• Two years until characterized as auctions based on received bandwidth

•Lots of incremental research!• Every 10-line tweak => 1 paper

• Hacks: Share ratios, manual pruning, …

• We lack the tools to tackle difficult problems

Antfarm Goals•High performance•Low cost of deployment•Performance guarantees• Administrator control over swarm

performance

•Accounting• Enables different resource

contribution policies

Antfarm Approach

•Key insight: view content distribution as an optimization problem

•Hybrid architecture

• P2P swarming with a logically centralized coordinator

•Clean slate protocol

Antfarm System Model

coordinatorseeder

altruist

Strawman Coordinator

•One could schedule every data transfer in the system

• All packets for all time

• Unscalable, impractical!

•Antfarm coordinator makes critical decisions based on observed dynamics

Antfarm Coordinator

•Models swarm dynamics• Measures and extracts key

parameters

•Formulates optimization problem• Calculates optimal bandwidth

allocation

•Enacts allocation decisions• Maximizes aggregate bandwidth

• Minimizes average download time

Antfarm Formalization

Maximize system-wide aggregate bandwidth subject to a bandwidth constraint

Response Curves

slope = 1 slope = 0

Response CurvesSwarm aggregate bandwidth (KB/s)

1500

1000

500

0

Seeder bandwidth (KB/s)0 25 75 10050

Swarm Dynamics

Swarms exhibit different dynamics based on size,

peer resources,network conditions. . .

Antfarm Optimization

σA σCσB

σA + σB + σC = B

A

B

C

Performance Control

•Can provide swarm performance guarantees

• Guarantee minimum level of service

• Prioritize swarms

Antfarm Allocation

σA′ σC′σB′

σA′+ σB′+ σC′= B

A

B

C

Problems•The optimization problem has limitations

•Requires knowing total bandwidth constraint

•Applies only to full caches under the control of the coordinator

•At the wire protocol level, there are many opportunities for malfeasance

Antfarm++•Use a currency scheme, called karma, to keep track of peers’ resource consumption and contribution

•Every exchange costs some karma to the peer receiving a block, brings karma to the peer providing the block and bandwidth

peer A

purseledger

Wire Protocol

•Coordinator mints small, unforgeable tokens

•Peers trade each other tokens for blocks

•Peers return spent tokens to the coordinator as proof of contribution

coordinator

peer B

purseledger

Tokens & Exchange Rates

•Every swarm uses a per-swarm currency, every token has source-destinations marked

•Token flows reveal the internal network dynamics to the coordinator

•The coordinator can provide incentives in the form of exchange rates

•Exchange rates encode the differential in the desirability of participating in different swarms => no need to estimate total BW

Benefits of a Virtual Currency•The cost and benefit of a transaction need not be symmetrical

•A block always costs 1 token to the buyer

•A block can bring γii karma to the seller

•Provide an incentive for peers to Provide an incentive for peers to join and seed swarms in needjoin and seed swarms in need

Antfarm Performance

0

500

1000

1500

2000

2500

3000

3500

4000

4500

Zipf, 60 KB/s seeder Zipf, 200 KB/s seeder

Aggr

egat

e ba

ndwid

th (K

B/s)

Client-server BitTorrent Antfarm

Swarm Starvation

BitTorrent starves the singleton swarm

BitTorrent: Starves New Swarm

Swarms, ordered largest to smallest

newself-

sufficientsingleto

n

Antfarm: Seeds New Swarm

Swarms, ordered largest to smallest

newself-

sufficientsingleto

n

Conclusions

•Hard problems in systems research can benefit immensely from the application of economics

• http://flixq.com showcases our current work

Related Work•Content Distribution Networks

• Akamai, CoBlitz, CoDeeN, ECHOS, Coral, Slurpie, YouTube, Hulu, GridCast, Tribler, Joost, Huang et al. 2008, ...

•P2P Swarming

• BitTorrent, BitTyrant, PropShare, BitTornado, BASS, Annapureddy et al. 2007, Guo et al. 2005, ...

•Incentives and microcurrencies

• Dandelion, BAR Gossip, Samsara, Karma, SHARP, PPay, Kash et al. 2007, ...

Questions?

PlanetLab Deployment

PlanetLab Deployment

0

2375

4750

7125

9500

11875

50 100 200

Seeder bandwidth (KB/s)

Aggre

gate

bandwid

th (KB/s

)

Client-server BitTorrent Antfarm