Chord-over-Chord Overlay

Sudhindra Rao

Ph.D Qualifier Exam

Department of ECECS

Outline

Peer-to-Peer systems Centralized P2P systems (hybrid) Unstructured P2P systems (pure) Structured P2P systems Super-peer networks CoCO Analysis and Conclusion

Peer-To-Peer Systems Decentralized data and resource sharing All computers have equal capabilities The resources can include:

Processing power Data Network bandwidth

Applications Redundant storage Permanence Selection of nearby servers Anonymity Search Authentication Hierarchical naming

Centralized Server P2P systems - Napster

Used in large scale sharing of files Single server maintains a table of

data Vs node Features

Not self-organized Not scalable Single point of failure/attack Most popular network - mp3

sharing Applications:

Napster

Central Server

Peer 5

Peer 2

Peer 3

Peer 4

Unstructured P2P networks - Gnutella

Random overlay networks No central index Start with nodes that know

about peer servers and flood along the network

Peers find neighbors Features:

Scalability – Flooding limited by TTL

Keyword search Cannot guarantee

search Applications:

Gnutella Freenet

Peer 1

Peer 6 Peer 7

Peer 4

Peer 3

Peer 5

Peer 2

Structured P2P networks – Chord, Pastry, CAN

Based on ‘Distributed Hash Tables’

Self-organized overlay networks

Insertion and lookup in a bounded number of hops

Features: Load balancing Fault-tolerance Decentralization Scalability Availability Flexible naming

Applications: Chord Pastry Tapestry CAN

N51

N8

N14

N21

N1

N38N42

N48

N58K54

Lookup (K54)

N32

Design and Analysis Chord provides fast distributed computation of a

hash function, mapping keys to nodes responsible for them

Assigns keys to nodes with consistent hashing A chord node needs only a small amount of routing

information about other nodes A node resolves the hash function by

communicating with other nodes With high probability, the number of nodes that

must be contacted to find a successor is an N-node network is O(log N)

Only O(log N) fingers need be stored When an Nth node joins or leaves the network,

only an O(1/N) fraction of the keys are moved

Super-Peer Networks Hierarchy introduces manageability Super-Peer networks combine features of distributed

search and centralized search Super-Peer node acts as server for subset of peers Inherent heterogeneity in the capability of peers on

the network Super-Peers are assigned based on processing power,

network bandwidth, degree etc. Super-peers communicate by flooding to other super-

peers Super-peer to peer communication – centralized

server system

Super-peer network Thumb rules for design

Increasing cluster size reduces aggregate load

Super-peer redundancy makes system resilient

Super-peers should have higher out-degree

Minimize TTL on floods Drawbacks

Flooding does not guarantee search success

Super-peers can be burdened

Flooding traffic and duplicates

Self-similarities in Mandelbrot Set

Central Server0 and 640

Super Peer128 - 255

Super Peer256 - 383

Super Peer384 - 511

Super Peer512 - 639

Super Peer1 - 127

Chord-over-Chord Overlay

Chord-over-Chord Overlay(CoCO)

Chord used in local clusters – Super-peer as manager

Super-peer redundancy - by assigning super-peers at the edge

Super-peers form a Chord overlay network

Super-peers maintain finger tables for cluster as well as the super-peer overlay

Central Server consulted only if all Chord searches fail on the overlay

Central Server

Super Peer

Super Peer

Super Peer

Super PeerSuper Peer

Chord-over-Centralized Server Overlay

Chord-over-Centralized server Overlay (CoCO)

Super-peers maintain a direct link to the Central Server

Central Server consulted in case of failed searches in local clusters

Central Server may be single point of failure

CoCO Analysis

Number of nodes to be contacted in the local cluster of size N/m - O(log N/m)

Cost of searching on Super-peer overlay - O(log m) Only O(log N/m) fingers need to be stored in peers

and O(log m) additional fingers on super-peers When an node joins or leaves the network, only an

O(m/N) fraction of the keys are moved and when Super peer leaves a network chord flip reassigns O(log N/m) + O(log m) fingers.

Discussion CoCO

Uses DHT on all layers – hence resilient to failures, attacks.

Increasing hierarchy improves manageability like Internet

Efficient and guaranteed search results Joins/Leaves handled efficiently Super-Peer reassignment is integral part of the protocol

Super-Peer networks using Gnutella Flooding can reduce efficiency Techniques to reduce flooding directly affect the

network efficiency Super-peer failures are not accounted for Flooding on super-peers does not guarantee search

results

Conclusion Possible applications of CoCO

University wide P2P networks Each department has its own super-peer

Company wide P2P networks Geographically distant networks controlled by

administrators – super-peer assignment ISP controlled Napster like central server

Strategically placed Super-peers – like Akamai caches

Better control over the network dynamics and easy to implement

Structured network is key to simpler administration

Thank you!