Rarest First and Choke Algorithms Are Enough speaker: Yongming Chen 2007.05.10.

Rarest First and Choke Algorithms Are Enough

speaker: Yongming Chen

2007.05.10

Outline

• Introduction

• Methodology

• Experimental Results

• Conclusion

Introduction

• recent studies have proposed replacement of rarest first and choke algorithm to improve BT

• Didn’t Cohen think of efficiency and fairness?

Terminology

• pieces and blocks• interested and choked• peer set• leecher and seed

– peer state

• rare and available pieces– # of copies

• rarest pieces– least # of copies

BitTorrent

• Rarest First Algorithm– piece selection strategy– download at random in rarest pieces set– random first, strict priority, end game mode

• Choke Algorithm– peer selection strategy– regular unchoked peers and optimistic unchoked peer

in leecher state– seed kept unchoked peers and seed random unchoke

d peer in seed state

BitTorrent

• rarest first may lead to the scarcity of some pieces

• choke algorithm is not fair and favors free riders

Experiment

• instrument a mainline client– add some logs

• default parameters: – 20kB/s max. upload rate, no download limit– max. peer set size is 80– 4 unchoked peers

• limitations: – single client instrumentation– limited torrent set

Torrents

Methodology

• use entropy instead of system capacity to evaluate efficiency of a piece selection strategy– characterized with peer availability

• peer availability of x according to y = (the time peer y is interested in x) ∕ (the time x is in the peer set of y)

Methodology

• bit level tit-for-tat fairness is not appropriate in the context of p2p file replication– excess capacity will be lost– a seed cannot evaluate the reciprocation of a leecher

• alternative:

– for any leecher i, j, Di < Dj if Ui < Uj

– a seed should give the same service time to each leecher

Results

Rarest First Algorithm

• almost all torrents have the ideal entropy– 1,2,4,5,6,7,8,9 in top graph are in startup pha

se– 2,18,19,21,26 in bottom graph have a small #

of leechers

• rarest first does not perform optimally in some extreme cases– peer join with only a few missing pieces– super seeding

Results

Results

Rarest First Algorithm

• transient state: only one seed

steady state: no rare pieces

• the duration of transient phase is constrained by the initial upload capacity– rarest first attempts to minimize it

• rarest first algorithm is efficient at keeping a torrent in steady state

Results

Results

Choke Algorithm

• strong correlation between the amount of bytes uploaded and downloaded

• choke algorithm leads to an equilibrium in the peer selection

• new version of choke algorithm in seed state is a significant improvement over the old one

Conclusion

• The rarest first algorithm guarantees a high diversity of the pieces.

• first block problem rather than last pieces problem

• To achieve fairness with a bit level tit-for-tat strategy is not appropriate

• The choke algorithm is fair, fosters reciprocation, and is robust to free riders

Conclusion

• piece and peer selection are key ideas for designing a scalable p2p system

• choke is simple but not easy to explain

Reference

• A. Legout, G. Urvoy-Keller, and P.Michiardi. Rarest first and choke algorithms are enough. In Proc. ACM SIGCOMM’06

• http://wiki.theory.org/BitTorrentSpecification• S. Jun and M. Ahamad. Incentives in bittorrent in

duce free riding. In Proc. SIGCOMM’05 Workshops