evolution and enhancement of bittorrent network topologies
DESCRIPTION
Evolution and Enhancement of BitTorrent Network Topologies. authors: Cameron Dale, Jiangchuan Liu, Joseph Peters, Bo Li presented by: Cameron Dale Simon Fraser University Burnaby, BC, Canada [email protected] IWQoS, June 2 nd , 2008, University of Twente, Enschede, The Netherlands. - PowerPoint PPT PresentationTRANSCRIPT
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Evolution and Enhancement of BitTorrent Network Topologies
authors:
Cameron Dale, Jiangchuan Liu, Joseph Peters, Bo Li
presented by:
Cameron DaleSimon Fraser University
Burnaby, BC, [email protected]
IWQoS, June 2nd, 2008, University of Twente, Enschede, The Netherlands
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Overview
● Network topology characteristics
● BitTorrent and its networks
● Experiments and Results
● Enhancement and Results
● Discussion
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Scale-Free Networks
• node degree k has a power law distribution
• characterized by:– evolving networks– preferential attachment
• tolerant of random node failures
kkP )(
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Small-World Networks
● characteristics of both random and regular networks
● large amounts of clustering of nodes– Clustering Coefficient (CC): the fraction of all
possible edges that exist between the neighbors of a node
● short average distance between nodes– Characteristic Path Length (CPL)
● known to be effective for the exchange and dissemination of information
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Motivation
● scale-free networks are resilient
● small-world networks are efficient
● both are desirable for file sharing, and have been found in other P2P networks
● recent papers have found clustering in some of the BitTorrent networks
● all previous results have examined only the initial stage of BitTorrent networks
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
BitTorrent
● peer-to-peer file sharing program
● peer connects to a tracker to find other peers– tracker returns random list of 50 peers
● peer connects to other peers– number of connections is usually limited– default value for most clients is 80
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
BitTorrent Networks
● Connection: the neighbors of a peer, to which it is connected by TCP (undirected)
● Interest: a peer is interested in other peers that have pieces of the download they need
● Unchoked: a peer's indication of a willingness to upload to another peer (incentive mechanism)
● Download: a peer is downloading from other peers
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Related Work
● Al-Hamra et. al. examined the diameter and found clustering in the Connection network
● Legout et. al. found clustering in the Unchoked network
● both used qualitative methods to demonstrate the clustering
● both examined only the very early stage of a BitTorrent download
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Basic Experimental Design
● modified BitTornado client
● run on more than 400 PlanetLab nodes
● download a 780 MB sample file
● node will connect, download, stop, delete, and reconnect multiple times
● experiments run for over 100 hours
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Other Experiments
● remove the limit on the number of connections a peer can make (was 80)– no effect on most of the results– connectivity matrix is random throughout the
experiment
● peers join and leave in groups, causing periods of increased and decreased churn– only a small effect on the Clustering Coefficient
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Distribution of Peer Characteristics
Nodes
45% 44 hours 6 kB/s 6 kB/s 2
25% 20 hours 12 kB/s 24 kB/s 3
15% 12 hours 25 kB/s 75 kB/s 4
10% 6 hours 50 kB/s 150 kB/s 5
5% 4 hours 100 kB/s 500 kB/s 6
Restarts After
Upload Speed
Download Speed
Unchoke Slots
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Peer Population
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Scale-Free Analysis
● plot the degree of each node versus its rank on a log-log scale
● linear fit gives the power law exponent γ and an R2 goodness of fit value
● only the Unchoked network had high R2 values of 0.9 (others were less than 0.7)
● previous real-world experiments show power law exponents between 2 and 3
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Sample Node Degree Fit
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Power Law Exponent
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Small-World Analysis
● calculate the Characteristic Path Length and Clustering Coefficient of each network at periodic intervals
● also generate a similar-sized random graph for comparison
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Characteristic Path Length
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Characteristic Path Length
● varies only during the initial stage
● short due to the density of the graph– 440 nodes with average degree 65
● very close to the random graph's CPL
● Unchoked graph is slightly larger– may be due to scale-free nature of the graph
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Clustering Coefficient
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
CC Compared to Random Graph
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Clustering Coefficient
● most of the variation is in the initial stage, and none is seen after the transient stage
● Unchoked network shows some clustering in the initial stage
● other than that, there is very little clustering, and none after the transient stage
● lack of clustering means no small-world– most networks are essentially random
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Connectivity Matrix
● scatter plot of peer connections– point at (x, y) indicates that peer x is connected
to peer y
● peers are assigned an index based on the time they joined the system
● early peers maximize their number of connections quickly
● used by Al-Hamra et. al. to indicate clustering in the Connection network
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Connectivity Matrix Comparison
Our Results (Hour 4) Al-Hamra et. al.
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Connectivity Matrix Evolution
Hour 4
Hour 8
Hour 16
Hour 32
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Connectivity Matrix
● results for initial stage are very similar to previous results
● later results show increased randomness
● in the end, the connectivity matrix is completely random
● the experiment with no limit on the number of connections a peer can make showed a completely random connectivity matrix throughout the experiment
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Adding Small-World to BitTorrent
● design a tracker that creates a small-world BitTorrent network of peers– only realistic modification possible for BitTorrent
● try similar strategy to Watts and Strogatz– start with a regular graph with a maximized
Clustering Coefficient– add some randomness
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
A Cycle of n-cliques
• connect n-cliques in a cycle• gives a Clustering Coefficient of
• independent of the total number of nodes• for BitTorrent, n = 80, so CC = 0.9986• increases the CPL to twice that of a random
graph (3.8)
)1(
61
nnCC
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Finding the Small-World
● add some randomness to the regular graph
● replace an n-clique edge with a random edge that connects together n-cliques
● vary the amount of randomness (number of replaced edges)
● graphs usually become small-world with a randomness of 1% to 3%
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Adding Randomness to n-cliques
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Small-World Tracker
● assign each new peer to an unfilled n-clique– if all are full, create a new n-clique
● choose the list of peers to return:– a small fixed number from other n-cliques– the rest (the majority) from the peer's n-clique
● two configuration parameters– random-peers: the small fixed number chosen
from other n-cliques– clique-size: the maximum size of an n-clique
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Simulation (clique-size = 80)
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Simulation (N = 400)
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Small-World Tracker Experiment
● experiment is mostly unmodified from the previous one
● only change is the Small-World tracker– random-peers: set to 1– clique-size: set to 40
● same results for peer population, Unchoked network is still scale-free
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Small-World Tracker CPL
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Original CC (different scale)
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Small-World Tracker CC
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Original CC Compared to Random
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Small-World Tracker CC to Random
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Small-World Connectivity Matrix
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Small-World Tracker Results
● small-world tracker has substantially increased the clustering of peers– all networks were affected, though our changes
were only to the Connected network
● the Characteristic Path Length also increased by 10% to 20%
● the BitTorrent peers now form a small-world network
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Initial Stage Not Typical
● long-term experiments are needed
● caused by peer dynamics over time– peers become seeds and drop connections to
other seeds– leaving peers free up connection slots to others– new peers are requested periodically from the
tracker
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Scale-Free Unchoked Network
● power law distribution of nodes with exponent of 2
● unaffected by churn and other changes
● may explain some of BitTorrent's efficiency– scale-free graphs are resistant to churn
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Lack of Clustering
● almost no clustering in any of the networks
● almost all graphs are completely random– not surprising due to the random peer lists
returned by the tracker
● therefore, no small-world networks are possible
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Introducing the Small-World
● theoretical framework for creating dense small-world graphs
● simple implementation in a BitTorrent tracker
● dramatically increases the clustering of peers
● all networks show small-world properties
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Future Work
● experimentally determine the efficiency of small-world BitTorrent networks
● improvements to the Small-World Tracker:– consider effect of small n-cliques on peers– what happens when many peers leave
(consolidating n-cliques)
● further examination of the effect of churn on the clustering of BitTorrent peers
Introduction Experiments Results Enhancement Results Discussion
Cameron Dale Evolution and Enhancement of BitTorrent Network Topologies
Evolution and Enhancement of BitTorrent Network Topologies
authors:
Cameron Dale, Jiangchuan Liu, Joseph Peters, Bo Li
presented by:
Cameron DaleSimon Fraser University
Burnaby, BC, [email protected]
IWQoS, June 2nd, 2008, University of Twente, Enschede, The Netherlands