Scribd
Upload
114 views23 pages

Clustering and Sharing Incentives in Bittorrent Systems

This document summarizes a study on the impact of initial seed upload capacity on peer clustering, sharing incentives, and upload utilization in BitTorrent systems. The study found that a well-provisioned initial seed leads to clear peer clustering by class, effective sharing incentives where faster upload speeds result in faster download speeds, and high upload utilization. However, an underprovisioned initial seed results in no discernible peer clustering, ineffective sharing incentives where download speeds are not correlated to upload speeds, and potentially reduced but still high upload utilization. The conclusion is that the initial seed's upload capacity is critical to the effectiveness of BitTorrent's choking algorithm and robustness during startup.

Uploaded by

sushmsn
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPT, PDF, TXT or read online on Scribd
114 views23 pages

Clustering and Sharing Incentives in Bittorrent Systems

This document summarizes a study on the impact of initial seed upload capacity on peer clustering, sharing incentives, and upload utilization in BitTorrent systems. The study found that a well-provisioned initial seed leads to clear peer clustering by class, effective sharing incentives where faster upload speeds result in faster download speeds, and high upload utilization. However, an underprovisioned initial seed results in no discernible peer clustering, ineffective sharing incentives where download speeds are not correlated to upload speeds, and potentially reduced but still high upload utilization. The conclusion is that the initial seed's upload capacity is critical to the effectiveness of BitTorrent's choking algorithm and robustness during startup.

Uploaded by

sushmsn
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPT, PDF, TXT or read online on Scribd
You are on page 1/ 23

Clustering and Sharing Incentives

in
BitTorrent Systems

Arnaud Legout1, Nikitas Liogkas2, Eddie Kohler2, Lixia Zhang2

1
INRIA, Projet Planète, Sophia Antipolis, France
2
UCLA, Los Angeles, CA USA
BitTorrent Overview
coolContent.torrent Web server
Which
random peers to ask pieces from?
peer set
Which piece to download first?
Tracker

coolContent.xvid

P1 P2 P3

2
Peer and Piece Selection
 Peer Selection
 Choking algorithm
 Maximize performance
 Foster reciprocation and
penalize free riders
 Piece selection
 Rarest first algorithm
 Enable high piece diversity
• Always find an interesting
piece at any other peer
• Does not bias peer
selection
 Focus on peer selection
3
Choking Algorithm
 Choking algorithm
 Local and remote peers
 Choke and unchoke
 Leechers: upload to the peers
(regular unchoke) from which
we are downloading the
fastest
• Re-evaluate periodically (10s)
 Optimistic unchoke
• Changed periodically (30s)
 Typically: 3 regular unchokes
+ 1 optimistic unchoke
 Seeds: refer to the paper
4
Choking Algorithm Properties
Does the choking algorithm
 Converge to an equilibrium?
• Speed and stability
 Provide effective sharing incentives?
• How much do I gain if I contribute
 Reach optimal efficiency?
• How far is it from a 100% upload utilization

What is the impact of the initial seed upload


capacity on those properties ?
5
Outline
Background and Motivation
Methodology
Results
 Fast Seed
 Slow Seed
Conclusion

6
Methodology: Experiments
Instrumentation of around 40 peers on
PlanetLab
 1 single initial seed connected for the
duration of experiment
 40 leechers join at the same time (flash
crowd) and leave as soon as they have the
content
 All peers (seed + leechers) use an
instrumented client
 Content: 113MB, 453 pieces (256kB each)
7
Methodology: Experiments
Leechers upload limit
 Umax is the maximum upload speed
 Three-class scenario
• 13 slow leechers with Umax= 20kB/s
• 14 medium leechers with Umax= 50kB/s
• 13 fast leechers with Umax= 200kB/s
Seed upload limit
 Three types of experiments
• 200 kB/s, 100 kB/s, and 20 kB/s
No download limitation for leechers 8
Metrics
Clustering index (cluster creation)
 Convergence?
Completion time (sharing incentives)
 How does a peer’s upload speed affects its
download speed?
Upload utilization (efficiency)
 What is the peers’ upload utilization?

9
Outline
Background and Motivation
Methodology
Results
 Fast Seed
 Slow Seed
Conclusion

10
Peer Clustering: Fast Seed
 Three-class scenario,
averaged over all 13 runs
 Seed max upload speed:
200kB/s
 We see clusters per class
 Two artifacts
19
 Slow class squares are
darker since peers take
longer to complete
 Peer 27 slower than other
peers in its class (problem
with a PlanetLab node):
20 Reciprocates mainly with the
slow medium fast slow leechers
Peer 27 seed
11
Clustering Index
 Clustering index of a peer P for class C

 duration of regular unchokes of P to i


ipeers in C
IC (P ) 
 duration of regular unchokes of P to i
i all peers

 IC(P)=1 if P unchoked only peers in class C


 IC(P)=0 if P unchoked only peers not in class C
 IC(P)=0.3 if P unchoked peers uniformly at
random (with 3 classes) 12
Peer Clustering: Fast Seed
 Three-class scenario,
averaged over all 13
runs
90th  Seed max upload speed:
200kB/s
10th
 Each peer has a high
clustering index to
peers in its class
 Peers of a specific class
prefer to unchoke peers
in the same class
7

Clusters of peers in the same class 13


Sharing Incentive: Fast Seed
 Three-class scenario, for
all 13 runs
 Seed max upload speed:
200kB/s
 Fast peers complete close
to earliest possible
completion time
 The more you contribute
the faster you complete

Earliest possible completion


Effective sharing incentive
time
14
Upload Utilization: Fast Seed
 Three-class scenario, for
all 13 runs
 Seed max upload speed:
200kB/s
 Each dot is the average
upload utilization over all
peers for a single run
 Upload utilization close to 1
 Room for improvement at the
beginning (bandwidth
matching to improve cluster
formation)

High upload utilization 15


Outline
Background and Motivation
Methodology
Results
 Fast Seed
 Slow Seed
Conclusion

16
Peer Clustering: Slow Seed

No discernible clusters

seed
 Three-class scenario, averaged over all 8 runs
 Seed max upload speed: 100kB/s

Fast peers break all clusters


17
Sharing Incentive: Slow Seed
 Three-class scenario,
for all 8 runs
 Seed max upload speed:
100kB/s
 Most peers complete
close to the earliest
completion time
 Choking algorithm does
not provide effective
sharing incentive when
the seed is
underprovisioned
 Earliest completion time
longer than with a fast
Earliest possible completion seed
time
No effective sharing incentives 18
Upload Utilization: Slow Seed
 Three-class scenario,
for all 8 runs
 Seed max upload speed:
100kB/s
 Each dot is the average
upload utilization over
all peers for a single run
 Still fairly high upload
utilization
 With an initial seed at
20kB/s the upload
utilization falls to 0.2
Upload utilization depends on seed upload
19
speed
Outline
Background and Motivation
Methodology
Results
 Fast Seed
 Slow Seed
Conclusion

20
Conclusion
 Seed provisioning is critical to the choking
algorithm’s effectiveness
 Well-provisioned initial seed
 Cluster formation, effective sharing incentive,
good upload utilization
 Underprovisioned initial seed
 No clustering, ineffective sharing incentives,
upload utilization can still be high

 What is the practical impact of these results?


21
Conclusion
 It has been known that the initial seed upload speed
is critical to the service capacity of torrents in their
starting phase
 We have shown that it is also critical to the
robustness of the torrent to free riders

The seed should be at least as fast as the


fastest leechers to support a robust
 Rule oftorrent
thumb during the startup phase
 Depends on how many fast peers there are
 Would be nice to model

22
Clustering and Sharing Incentives in BitTorrent Systems
Arnaud Legout, Nikitas Liogkas, Eddie Kohler, Lixia Zhang

Questions?

Instrumented client available at:


http://www-sop.inria.fr/planete/Arnaud.Legout/Projects/p2p_cd.html

You might also like