Simulating Gnutella 0.6

Communication NetworksWei Lu, Sireen Malik

Simulating Gnutella 0.6

Student Project Final PresentationWei Lu 02.2005

-2-Communication NetworksWei Lu, Sireen Malik

Gnutella• Gnutella 0.4

– Distributed Index Flooding Architecture (DIFA)

– Signaling• Attempt connection with Bootstra

p servers• Once connected, flood PINGS• Reverse routed PONGS from kno

wn-hosts• PONG-based connection to know

n hosts• Flood QUERIES to all known-ho

sts• Reverse routed QUERYHITS

– File download via HTTP-TCP[Source:P2P Traffic Simulation, Sireen Malik, Prof. Ulrich Killat]

• Gnutella 0.6– Two-layered hierarchy

• Supernodes (called Ultrapeers) • Client nodes (called Shielded Leave

s)• Ultrapeers: faster,better networking

and CPU power act as proxies for client(leaf) nodes connected to them shield leaf nodes from almost all ping and query traffic


Gnutella 0.6, Anti-flooding

• Message Dispatching– Leaf node doesn't relay Ping message– Ultrapeer only dispatches Ping message to

other connected Ultrapeers.• Query Routing

– Leaf node doesn't spread Query message.– Ultrapeer only dispatches Query message to oth

er Ultrapeer and some of his leaf nodes selectively (DIHA).


Gnutella 0.6, Ultrapeer Election• New header fields for handshaking

– X-Ultrapeer, whether a host plans on acting as ultrapeer– X-Ultrapeer-Needed, to balance the number of ultrapeers

• Handshaking– Leaf to Ultrapeer, Leaf to Shielded Leaf, Leaf to Unshielded Leaf,

Ultrapeer to Ultrapeer

• Specification is quite open on details– Max-number of leaves (Cluster Size): LimeWire 30, BearShare 45.– Ultrapeer critera: “at least 15KB/s downstream and 10KB/s upstre

am”, “have been running for few hours/sever minutes”, “when (not) too many”…

• Ultrapeer : ?%, Leaf: ?%


Optimal Topology

• What’s the globally optimal topology? (% of UltrapeerCluster size, Degree distribution, etc.)

• How to locally approach the optimal topology?(Ultrapeer criteria, Routing protocol)


Optimal Topology, Q1• Analytic model: not well established yet.

– Parameters• Ultrapeer percentage, Cluster Size, Degree distribution Pd, TTL, Q

uery rate Rq, Replication rate Rr, …– Objectives

• Load (Computation, Communication), Quality of result, …– Random Graph, Load{QueryTraffic[QueryCoverage(Pd, T

TL)], QueryHitTraffic(Rr, Rq)}. Hard to calculate.[1]– Var(load(p)/CHL(p)), recursive function, simulated annealin

g to calculate.[2]• Measurement: strongly related to software configurat

ion/ protocol implementation.


Optimal Topology, Q2

• [3] gives rules of thumb.– But the model based on semi-p2p system (superp

eer (server) is at most 1 hop away) might not be accurate enough.

• Cluster head selection in ad-hoc and sensor network[4][5]. – Based on Node ID, Node degree. Optimize pow

er consumption (load), communication, etc.


Optimal Topology, dynamic network

• Most of above works are based on static context.

• Peer availability matters.


Current Approach

• Related to Implementation[Daniel Stutzbach, Reza Rejaie, “Characterizing Today’s Gnutella Topology”, Dec.2004]

• Our config:5% Ultrapeer.


Implementation, NS-2

GnutellaServentGnutellaApp

TclObject

GnutellaGlobalRegistrationOffice

GnutellaConnection

TcpAgent

-tcp_send,tcp_sink,dst_send,dst_sink 1*

-connection_list_

1

*

-gnuProt_

1

-gnutella_appl_

*

Node

-node_

1

-node_ 1

-app_ 1

• Extending Gnutella 0.4: Ultrapeer Election, DIHA, Downloading, ...

• Facilities: control of selected peers, trace on link, random seed, ...


Implementation, Ptolemy

• Ptolemy– Adapting NS-2 code to Ptolemy.– No modification, just re-compile.

Real World NS-2 Ptolemy

Network Interface Node Star / Galaxy

TCP Agent Star / Galaxy

Gnutella Application Star / Galaxy

Adapter

Adapter


Configuration – User Model• All MODEM & DSL users in ON/OFF process

– mean on 60 sec, log quadratic; mean off 105 sec, neg. exp.

• Every 25s, drops oldest connections when many• File size distribution:

– for < 10MB files (music) less than 15% bytes but more than 88% request

– for >100MB files (movies) more than 65% bytes but less than 5% request

• Number of files shared by peers– 66% free-riders, 73% share 10 or less files, top 1%

share 35% files


Configuration – Network• Germany Network, 17

core nodes.• Power Law for 2nd level

node connectivity (Pareto Distribution, Shape 1.5)

• Bandwidth of 2nd level nodes (70% MODEM, Uniform Distribution)


Result – Average Signaling Traffic MODEM (100Kbps)


Result – Average Singling Traffic DSL (1~4Mbps)


Result – Average Signaling Traffic Bootstrap (1000Mbps)


Result – Signaling Sum-up• MODEM(100Kbps):

70%, DSL(1~4Mbps): 29.6% , Bootstrap(1000Mbps) : 0.4%.

• 35>1/5% (5% peers behave as in 0.4)– Less peers flood

message.– Message flooded in

smaller scope35)(

6.0

4.0 MsgRateMsgRateMean


Downloading• Previous configurations, e.g. File size distribution:

• Each QueryHit triggers a HTTP session for downloading.

• Result: QueryHit/Query = 34.73%


Result – Gnutella 0.6 Signaling and Signaling+Downloading

• By single MODEM(100Kbps) and DSL(1Mbps) user

• Signaling is minor part.• Log-Log scaled • MODEM:

– Synthesis ≈ 100 Signaling

• DSL:– Synthesis ≈ 1000 Signaling


References• [1] Ruediger Schollmeier, Gero Schollmeier, An analytic

model for the behavior of arbitary peer-to-peer networks.• [2] Mudhakar Srivatsa, Bugra Gedik, Ling Liu, Improving

Peer to Peer Search With Multi-tier Capability-Aware Overlay Topologies.

• [3] Beverly Yang, Hector Garcia-Molina, Designing a Super-Peer Network

• [4] Mario Gerla,Jack Tzu-Chieh Tsai, Multicluster, mobile, multimedia radio network

• [5] Ossama Younis, Sonia Fahmy, Distributed Clustering in Ad-hoc Sensor Networks: A Hybrid, Energy-Efficient Approach

Documents

Simulating Gnutella 0.6