university of jyvÄskylÄ topology management in unstructured p2p networks distributed systems...

17
UNIVERSITY OF JYVÄSKYLÄ Topology Management in Unstructured P2P Networks Distributed Systems Research Seminar on 22.3.2007 Annemari Auvinen, Research Student [email protected] Department of Mathematical Information Technology http://www.mit.jyu.fi/ cheesefactory

Post on 22-Dec-2015

217 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: UNIVERSITY OF JYVÄSKYLÄ Topology Management in Unstructured P2P Networks Distributed Systems Research Seminar on 22.3.2007 Annemari Auvinen, Research Student

UNIVERSITY OF JYVÄSKYLÄ

Topology Management in Unstructured P2P Networks

Distributed Systems Research Seminar on 22.3.2007

Annemari Auvinen, Research Student

[email protected] of Mathematical Information Technology

http://www.mit.jyu.fi/cheesefactory

Page 2: UNIVERSITY OF JYVÄSKYLÄ Topology Management in Unstructured P2P Networks Distributed Systems Research Seminar on 22.3.2007 Annemari Auvinen, Research Student

2007

UNIVERSITY OF JYVÄSKYLÄ

Content

• Topology Management • Topology Management Algorithms• Chedar and P2PRealm simulator• NeuroTopology• Future

Page 3: UNIVERSITY OF JYVÄSKYLÄ Topology Management in Unstructured P2P Networks Distributed Systems Research Seminar on 22.3.2007 Annemari Auvinen, Research Student

2007

UNIVERSITY OF JYVÄSKYLÄ

Topology Management 1/2

• Logical (i.e. overlay) topology on top of the physical network

• In an unstructured network a node's place in the network is not pre-defined like it is in a structured network

• A node may join the network by establishing a connection to another node on the P2P network

Page 4: UNIVERSITY OF JYVÄSKYLÄ Topology Management in Unstructured P2P Networks Distributed Systems Research Seminar on 22.3.2007 Annemari Auvinen, Research Student

2007

UNIVERSITY OF JYVÄSKYLÄ

Topology Management 2/2

• Topology management algorithms affect the topology by making network more scalable and effective for resource discovery

• Nodes are placed so that they stay connected and find resources efficiently without using too much of their capacity for being in the network

• Network can be kept connected• Self-organizing using local information

Page 5: UNIVERSITY OF JYVÄSKYLÄ Topology Management in Unstructured P2P Networks Distributed Systems Research Seminar on 22.3.2007 Annemari Auvinen, Research Student

2007

UNIVERSITY OF JYVÄSKYLÄ

Topology Management Algorithms

• Are based on the goodness of the node• A good neighbor node provides resources to

the node• Goodness is sum of

– The amount of the resource replies the node has got from the neighbor and

– The amount of the resource replies the neighbor’s neighbor has relayed to the node

Page 6: UNIVERSITY OF JYVÄSKYLÄ Topology Management in Unstructured P2P Networks Distributed Systems Research Seminar on 22.3.2007 Annemari Auvinen, Research Student

2007

UNIVERSITY OF JYVÄSKYLÄ

Node Selection and Node Removal

• Node searches a node to which to establish a new connection from the history based on hit values and request information

• Removed node is the ”worst” neighbor • Worst neighbor is a node which has the

lowest goodness value

Page 7: UNIVERSITY OF JYVÄSKYLÄ Topology Management in Unstructured P2P Networks Distributed Systems Research Seminar on 22.3.2007 Annemari Auvinen, Research Student

2007

UNIVERSITY OF JYVÄSKYLÄ

Overload Estimation

• Connections are established and dropped based on the traffic amount flowing through the node

• If the traffic meter value is more than the given traffic limit one node is dropped by using Node Removal

• If the traffic meter value is less than the given lower traffic limit, algorithm tries to establish a connection to new node by using Node Selection

Page 8: UNIVERSITY OF JYVÄSKYLÄ Topology Management in Unstructured P2P Networks Distributed Systems Research Seminar on 22.3.2007 Annemari Auvinen, Research Student

2007

UNIVERSITY OF JYVÄSKYLÄ

Overtaking

• Node moves closer to the ”good” nodes• If neighbor has neighbor whose relayed hits

proportion of all neighbor’s neighbors’ relayed hits and neighbor’s hits is more than the given percent node establishes a new connection to that node and current connection to the neighbor is dropped

1 23

4Hits:2

Relayed hits:6 (60%)

Relayed hits:2 (20%)

1 23

4

Page 9: UNIVERSITY OF JYVÄSKYLÄ Topology Management in Unstructured P2P Networks Distributed Systems Research Seminar on 22.3.2007 Annemari Auvinen, Research Student

2007

UNIVERSITY OF JYVÄSKYLÄ

Results

• Best combination of parameters: lower traffic limit 40%, 80% overtaking, traffic limit over 350 messages/50 sent messages

• Amount of changes in the network was small, topology got balance, neighbor distribution was power law and number of hops small

Page 10: UNIVERSITY OF JYVÄSKYLÄ Topology Management in Unstructured P2P Networks Distributed Systems Research Seminar on 22.3.2007 Annemari Auvinen, Research Student

2007

UNIVERSITY OF JYVÄSKYLÄ

Chedar

• Decentralized P2P middleware implemented using Java

• Basis for P2P applications: distributed computing (P2PDisCo), data fusion, extension for mobile devices

• Includes the topology management algorithms, but because of errors in connections and machines affect results -> algorithms were implemented and tested in P2PRealm simulator

Page 11: UNIVERSITY OF JYVÄSKYLÄ Topology Management in Unstructured P2P Networks Distributed Systems Research Seminar on 22.3.2007 Annemari Auvinen, Research Student

2007

UNIVERSITY OF JYVÄSKYLÄ

NeuroTopology

• Topology construction using neural networks• The idea is that every peer has a neural network to

make decisions about establishing new connections in a P2P network

• NeuroTopology algorithm is executed in every peer after a predefined amount of resource queries

• The algorithm goes through all neighbor candidates• The information that the neural network needs, is

gathered during resource queries

Page 12: UNIVERSITY OF JYVÄSKYLÄ Topology Management in Unstructured P2P Networks Distributed Systems Research Seminar on 22.3.2007 Annemari Auvinen, Research Student

2007

UNIVERSITY OF JYVÄSKYLÄ

Inputs

• Bias = 1.• CurrentNeighborsAmount is the number of node's neighbors • ToNeighborsAmounts is the number of node's candidate

neighbor’s neighbors • RepliesFromCandidates is the number of the resource replies

received from a candidate neighbor• RelayedRepliesFromCandidates is the number of the resource

replies which the candidate neighbor has relayed to the node• TrafficMeter is a counter, which calculates the amount of the

resource reply messages going through a node• TrafficLimit simulates the bandwidth of a candidate node. If

TrafficMeter value is bigger than Trafficlimit, the node will not reply to resource requests

Page 13: UNIVERSITY OF JYVÄSKYLÄ Topology Management in Unstructured P2P Networks Distributed Systems Research Seminar on 22.3.2007 Annemari Auvinen, Research Student

2007

UNIVERSITY OF JYVÄSKYLÄ

Training

• The weights of the neural network have to be optimized

• Evolutionary computing and Gaussian random variation were used

Define theP2P networkconditions

Define the fitness requirements

for the algorithm

Create candidate algorithmsrandomly

Select the bestones for next

generation

Breed a newpopulation

Finally select thebest algorithm forthese conditions

Iteratethousands

ofgenerations

Page 14: UNIVERSITY OF JYVÄSKYLÄ Topology Management in Unstructured P2P Networks Distributed Systems Research Seminar on 22.3.2007 Annemari Auvinen, Research Student

2007

UNIVERSITY OF JYVÄSKYLÄ

Fitness of the neural network

• The fitness function takes into account – The number of resource queries – The number of resource replies – The number of the packets the topology query

used – The number of the new connections, i.e. topology

changes

Page 15: UNIVERSITY OF JYVÄSKYLÄ Topology Management in Unstructured P2P Networks Distributed Systems Research Seminar on 22.3.2007 Annemari Auvinen, Research Student

2007

UNIVERSITY OF JYVÄSKYLÄ

ResultsA

lgor

ithm

Top

olog

y

Fitn

ess

Impr

ovem

ent

in F

itnes

s

Effi

cien

cy in

st

atic

to

polo

gy

Effi

cien

cy

Effi

cien

cy

(onl

y re

sour

ce

pack

ets)

HDS Grid 37502 24.76 % 0.145 0.207 0.236 RW Grid 34522 13.38 % 0.155 0.164 0.186 BFS Grid 22497 118.38 % 0.099 0.091 0.101 HDS Power 38130 -0.23 % 0.230 0.234 0.265 RW Power 37216 2.83 % 0.180 0.188 0.209 BFS Power 20768 13.53 % 0.097 0.085 0.093 HDS Random 37505 32.95 % 0.164 0.212 0.245 RW Random 35496 36.28 % 0.164 0.165 0.186 BFS Random 23382 20.90 % 0.140 0.095 0.106

Page 16: UNIVERSITY OF JYVÄSKYLÄ Topology Management in Unstructured P2P Networks Distributed Systems Research Seminar on 22.3.2007 Annemari Auvinen, Research Student

2007

UNIVERSITY OF JYVÄSKYLÄ

Future

• NeuroTopology: comparing the results with other algorithms

• Bayesian Networks and Petri Nets in Topology Management

Page 17: UNIVERSITY OF JYVÄSKYLÄ Topology Management in Unstructured P2P Networks Distributed Systems Research Seminar on 22.3.2007 Annemari Auvinen, Research Student

2007

UNIVERSITY OF JYVÄSKYLÄ

References

• Auvinen A., Vapa M., Weber M., Kotilainen N., Vuori J., "Chedar: Peer-to-Peer Middleware", Proceedings of the 19th IEEE International Parallel & Distributed Processing Symposium (IPDPS 2006), Rhodes Island, Greece, 2006.

• Kotilainen N., Vapa M., Keltanen T., Auvinen A., Vuori J., "P2PRealm - Peer-to-Peer Network Simulator", 11th International Workshop on Computer-Aided Modeling, Analysis and Design of Communication Links and Networks (CAMAD'06), IEEE Communications Society, pp. 93-99, Trento, Italy, 2006.

• Auvinen A., Vapa M., Weber M., Kotilainen N., Vuori J., ”New Topology Management Algorithms for Unstructured P2P Networks”, to be published in the Second International Conference on Internet and Web Applications and Services, May 2007.

• Auvinen A., Keltanen T., Vapa M., ”Topology Management in Unstructured P2P Networks Using Neural Networks”, submitted to IEEE Congress on Evolutionary Computation, March 2007.