antfarm: efficient content distribution with managed swarms ryan s. peterson, emin gun sirer usenix...

20
Antfarm: Efficient Content Distribution with Managed Swarms Ryan S. Peterson, Emin Gun Sirer USENIX NSDI 2009 Presented by: John Otto, Hongyu Gao 2009. 10. 21. Adapted from the slides of Eunsang Cho

Upload: joelle-peterkin

Post on 14-Dec-2015

216 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Antfarm: Efficient Content Distribution with Managed Swarms Ryan S. Peterson, Emin Gun Sirer USENIX NSDI 2009 Presented by: John Otto, Hongyu Gao 2009

Antfarm: Efficient Content Distribution with Managed Swarms

Ryan S. Peterson, Emin Gun SirerUSENIX NSDI 2009

Presented by:John Otto, Hongyu Gao

2009. 10. 21.Adapted from the slides of Eunsang Cho

Page 2: Antfarm: Efficient Content Distribution with Managed Swarms Ryan S. Peterson, Emin Gun Sirer USENIX NSDI 2009 Presented by: John Otto, Hongyu Gao 2009

Contents

• Problem Definition• Antfarm– Peer’s Perspective

– Coordinator’s Perspective

• Evaluation• Conclusion

2

Page 3: Antfarm: Efficient Content Distribution with Managed Swarms Ryan S. Peterson, Emin Gun Sirer USENIX NSDI 2009 Presented by: John Otto, Hongyu Gao 2009

Problem Definition

• To find an efficient way to disseminate a large set of files to a potentially very large set of clients

3

Page 4: Antfarm: Efficient Content Distribution with Managed Swarms Ryan S. Peterson, Emin Gun Sirer USENIX NSDI 2009 Presented by: John Otto, Hongyu Gao 2009

Existing Approaches

• Client-server– Pros: simple due to central authority

– Cons: cost and scalability

4

Page 5: Antfarm: Efficient Content Distribution with Managed Swarms Ryan S. Peterson, Emin Gun Sirer USENIX NSDI 2009 Presented by: John Otto, Hongyu Gao 2009

Existing Approaches

• Peer-to-peer swarms– Pros: reduced cost

– Cons: limited information, no control or performance guarantees

5

Page 6: Antfarm: Efficient Content Distribution with Managed Swarms Ryan S. Peterson, Emin Gun Sirer USENIX NSDI 2009 Presented by: John Otto, Hongyu Gao 2009

Goals

• High performance• Low cost of deployment• Performance guarantees– Administrator can control over swarm

performance

• Scalability

6

Page 7: Antfarm: Efficient Content Distribution with Managed Swarms Ryan S. Peterson, Emin Gun Sirer USENIX NSDI 2009 Presented by: John Otto, Hongyu Gao 2009

Antfarm

• Hybrid peer-to-peer architecture

• Content distribution optimization problem– Central authority (coordinator) makes decision

how to allocate bandwidth optimally.

7

Page 8: Antfarm: Efficient Content Distribution with Managed Swarms Ryan S. Peterson, Emin Gun Sirer USENIX NSDI 2009 Presented by: John Otto, Hongyu Gao 2009

System Overview

• Seeder: trusted servers managed by the coordinator that distribute data blocks to peers.

seeder

coordinator

swarm

8

Page 9: Antfarm: Efficient Content Distribution with Managed Swarms Ryan S. Peterson, Emin Gun Sirer USENIX NSDI 2009 Presented by: John Otto, Hongyu Gao 2009

Peer’s Perspective

• Default behavior– For peer and block selection is identical to

BitTorrent

• Advisory notification– Coordinator sends lists of underutilized peers as

candidates for data exchange.

• Token exchange– Incentive to data upload

9

Page 10: Antfarm: Efficient Content Distribution with Managed Swarms Ryan S. Peterson, Emin Gun Sirer USENIX NSDI 2009 Presented by: John Otto, Hongyu Gao 2009

Coordinator’s Perspective

• Coordinator– Collects statistics on peer network behavior

– Computes response curves and bandwidth allocations

– Steers the swarm toward an efficient operating point using token supply

• Formulation–Maximize system-wide aggregate bandwidth

subject to a bandwidth constraint

10

Page 11: Antfarm: Efficient Content Distribution with Managed Swarms Ryan S. Peterson, Emin Gun Sirer USENIX NSDI 2009 Presented by: John Otto, Hongyu Gao 2009

Constrained Optimization Problem

• Response curve– Critical properties of each swarm

– Primary input to the optimization problem

A: rapid increase

B: peer uplink capacity is exhausted

C: downlinks are saturated

11

Page 12: Antfarm: Efficient Content Distribution with Managed Swarms Ryan S. Peterson, Emin Gun Sirer USENIX NSDI 2009 Presented by: John Otto, Hongyu Gao 2009

Constrained Optimization Problem

• Coordinator “climbs” each of the curves, always preferring the steepest curve.

• E.g.) The optimal bandwidth allocation for three concurrent swarms.– All the allocation

points have the samederivative.

12

Page 13: Antfarm: Efficient Content Distribution with Managed Swarms Ryan S. Peterson, Emin Gun Sirer USENIX NSDI 2009 Presented by: John Otto, Hongyu Gao 2009

Performance Control and Adaptation

• Provides swarm performance guarantees– Guarantee minimum level of service

– Prioritize swarms

• Updates response curve–When swarm dynamics change

13

Page 14: Antfarm: Efficient Content Distribution with Managed Swarms Ryan S. Peterson, Emin Gun Sirer USENIX NSDI 2009 Presented by: John Otto, Hongyu Gao 2009

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.

14

coordinator

purse ledger

purse ledger

Peer A Peer B

Data block transfer

Page 15: Antfarm: Efficient Content Distribution with Managed Swarms Ryan S. Peterson, Emin Gun Sirer USENIX NSDI 2009 Presented by: John Otto, Hongyu Gao 2009

Performance Comparison

• Antfarm achieves the highest aggregate download bandwidth

15

Page 16: Antfarm: Efficient Content Distribution with Managed Swarms Ryan S. Peterson, Emin Gun Sirer USENIX NSDI 2009 Presented by: John Otto, Hongyu Gao 2009

Swarm Starvation

• Antfarm awards seeder bandwidth to the singleton swarm

16

Page 17: Antfarm: Efficient Content Distribution with Managed Swarms Ryan S. Peterson, Emin Gun Sirer USENIX NSDI 2009 Presented by: John Otto, Hongyu Gao 2009

New Swarm Starvation

• Antfarm achieves an order of magnitude increase in average download speed

17

Page 18: Antfarm: Efficient Content Distribution with Managed Swarms Ryan S. Peterson, Emin Gun Sirer USENIX NSDI 2009 Presented by: John Otto, Hongyu Gao 2009

PlanetLab Experiments

• Response curve

18

• Aggregate bandwidth

Page 19: Antfarm: Efficient Content Distribution with Managed Swarms Ryan S. Peterson, Emin Gun Sirer USENIX NSDI 2009 Presented by: John Otto, Hongyu Gao 2009

Scalability

• Even for large number of peers, the bandwidth consumption at the coordinator is modest.

19

Page 20: Antfarm: Efficient Content Distribution with Managed Swarms Ryan S. Peterson, Emin Gun Sirer USENIX NSDI 2009 Presented by: John Otto, Hongyu Gao 2009

Conclusion

• Antfarm models swarm dynamics and allocates bandwidth optimally.

• Novel hybrid architecture• Simulation and PlanetLab experiment show

that Antfarm outperforms client-server and BitTorrrent

20