LOOKING UP DATAIN P2P SYSTEMS

Hari Balakrishnan M. Frans Kaashoek David Karger Robert Morris

Ion StoicaMIT LCS

Key Idea

• Survey paper• Discusses how to access data in a P2P system• Covers four solutions

– CAN– Chord– Pastry– Tapestry

INTRODUCTION

• P2P systems are popular due to– Low startup cost– High scalability at very low cost– Use of resources that would otherwise remain

unused– Potential for greater robustness

• Fully decentralized and distributed

The lookup problem

• How do we locate data in large P2P systems?• One solution

– Distributed hash tables (DHT)

Previous solutions (I)

• Centralized database– Napster

• Not scalable • Vulnerable to attacks on database

Previous solutions (II)

• Broadcasting– Customers broadcast their requests to their

neighbors, which forward them to their own neighbors and so on

– Gnutella– Does not scale either

• Broadcast messages consume too much bandwidth

Previous solutions (III)

• Internet DNS– Organizes network nodes into an hierarchy– All searches start at top of hierarchy

• Propagate down– Used by KaZaA, Grokster and others– Nodes higher in the tree do much more work than

lower nodes– Solution vulnerable to loss of root node(s)

Previous solutions (IV)

• Freenet– Forwards queries from node to node until

requested data are found– Emphasis is on anonymity

• Not performance• Unpopular documents may become

inaccessible –Nobody cares!

DISTRIBUTED HASH TABLES

• Implements primitive lookup(key)– Produces a path going from a node no to the

node holding key

• Big tradeoff is between– Keeping paths short– Minimizing state information kept by nodes

Main design issues

• Mapping keys to nodes in a balanced way– Use a hash function

• Forwarding a lookup for a key to appropriate node– Find at each step a node closer to the node

holding the key• Building routing tables

– Each node should have a successor

CAN

• Uses a d-dimensional key space– Partitioned into hyper-rectangles

• "Zones"– Each node manages a zone

• Responsible for all keys in zone

Neighbors

• Each node keeps track of addresses of all its neighbors – Routing table

• Neighbors are defined as nodes sharing a (d-1) dimensional hyper-plane– Contacts with fewer dimensions in common

do not count

A two-dimensional example (I)

A two-dimensional example (II)

X(0, 0; 0.5, 0.5)

(0, 0) (1, 0)

(1, 1)(0, 1)

In reality the state space wraps

X(0, 0.5; 0.5, 1)

X(0.5, 0.5; 1, 1)

X(0.5, 0.25; 0.75, 0.5)

X(0.5, 0; 0.75, 0.25)

X(0.75, 0;1, 0.5)

A path from (0.25, 0.3) to (0.8, 0.8)

X(0, 0; 0.5, 0.5)

(0, 0) (1, 0)

(1, 1)(0, 1)

In reality the state space wraps

X(0, 0.5; 0.5, 1)

X(0.5, 0.5; 1, 1)

X(0.5, 0.25; 0.75, 0.5)

X(0.5, 0; 0.75, 0.25)

X(0.75, 0;1, 0.5)

Lookup

• Routing tries to approximate the straight path between current zone and zone holding the key

• Various optimizations attempt to reduce lookup latency

Dynamic behavior

• When a node joins the network – It picks random point in space– Find node managing the zone– Splits with it current zone

• When a node departs– Zones are merged

• More complex process

Fault-tolerance

• When a node fails neighbor with smallest zone takes over– Multiple failures may cause too many nodes

to handle multiple zones

CHORD

• Assigns ID's to keys and nodes in the same address space

• ID's are organized in a ring– ID 0 follows the highest ID

• Each node is responsible for all keys that immediately precede it in the key space

Example

N 4

K 6

N 12

N 20

N 24 K1

K 10K 15

Finger table

• Each node keeps a table containing IP addresses of nodes– Halfway around in the key space– Quarter-of-the-way around– …

• Table has log N entries– Allows O(log N) searches

Partial example

N 4

N 12

N 20

N 24

Fault-tolerance

• Each node has a successor list – Contains IP addresses of next r successors

• Guarantees routing progress as long as all r successors are not down

Dynamic behavior

• New node n learns its place in the Chord ring by asking any extant node to do a lookup(n)

• Must also– Update successor list of its predecessor– Create its own successor list

PASTRYPASTRY

• Scalable, self-organizing, routing and object Scalable, self-organizing, routing and object location infrastructurelocation infrastructure

• Each node has a node IDEach node has a node ID– IDs are uniformly distributed in the ID spaceIDs are uniformly distributed in the ID space

• Includes a Includes a proximity metricproximity metric to measure to measure distances between pairs of ID'sdistances between pairs of ID's

Pastry NodesPastry Nodes

• Each node maintains three sets of nodesEach node maintains three sets of nodes– Leaf setLeaf set

• Closest nodes in terms of node ID'sClosest nodes in terms of node ID's• Same function as Chord's successor listSame function as Chord's successor list

– Nodes in routing tableNodes in routing table• Prefix routing (big idea)Prefix routing (big idea)

– Neighborhood setNeighborhood set• Closest nodes in terms of proximity metricClosest nodes in terms of proximity metric

Dynamic behaviorDynamic behavior

• Pastry is self-organizingPastry is self-organizing– Nodes come and goNodes come and go– Includes a seed discovery protocolIncludes a seed discovery protocol

Prefix RoutingPrefix Routing

• At each step, a node forwards an incoming At each step, a node forwards an incoming request to a node whose node id has largest request to a node whose node id has largest common prefix with common prefix with

• Destination ID: Destination ID: 12301230• Node ID: Node ID: 1023023• Next Hop: Next Hop: 1212----

Routing table for node 1023

0221 2230 3120

1130 1233 1302

1003 1013 1032

1020 1022

No common prefixOne common digitTwo common digitsThree common digits

Routing request for node 1230

0221 2230 3120

1130 1223 1302

1003 1013 1032

1020 1022

No common prefixOne common digit

Two common digitsThree common digits

Request is always send to a node having at least one more common prefix digit. Here it's node 1223

At node 1233

0221 2230 3120

1030 1130 1302

1201 1211 1220

1230 1232

No common prefixOne common digitTwo common digitsThree common digits

Node with at least one more common prefix digitis node 1230

TAPESTRY

• Interprets keys as sequences of digits• Incremental prefix routing

– Similar to Pastry• Main contribution is emphasis on proximity

– In the actual world• Reduces query latency• Makes system much more complex

CONCLUSIONS

• Major issues include– Operational costs:

searches are all O(log n); storage costs vary– Fault-tolerance and concurrent changes:

only Chord and Tapestry can handle them– Proximity routing:

Pastry, CAN and Tapestry have heuristics– Malicious nodes:

Pastry checks node ID's

Summary of costs

CAN Chord Pastry Tapestry

Node state1 d log N log N log NLookup2 dN1/d log N log N log NJoin2 dN1/d +

d log Nlog2 N log2 N log2 N

1 number of other nodes known by a given