Smallworld Overlay P2P Network

1
Small-world Overlay P2P Network

• John C.S. Lui

2
Outline

• Describe our motivations on small world overlay
  P2P network.
• Introduce the background information of P2P
  network and small world network.
• Propose our Small-world Overlay Protocol (SWOP).
• Explain our Flash Crowd Handling Protocol
• Illustrate the experimental results

3
Motivation

• Fundamental Improvements in P2P network
• Improve the performance of Object Lookup in P2P
  Network
• Solve high traffic loading of a popular and
  dynamic object, i.e. under a Flash Crowd
  Scenario
• Small world is applied to achieve above criteria

4
Background--Structured P2P Network--

• P2P network contains nodes (computers), which are
  acting as server as well as client.
• Structured P2P with two extra characteristics
• Decentralized
• Structured
• To achieve this, consistent Distributed Hash
  Table (DHT) has been used.
• Two implementation issues for maintaining a
  logical structure
• Unique key assignment scheme using DHT
• Characteristic routing tableaims at reducing
  distance by at least half in each forwarding

5
Background--Structured P2P Network--
1. Unique key assignment
224
x.x.x.b
Item 1
x.x.x.c
x.x.x.a
Objects
Item 2
Item 3
x.x.x.d
Nodes
x.x.x.e
6
Background--Structured P2P Network--
2. Characteristic Routing
Item 1
Item 2
6000
1000
10000
12000
224
212-1
Item 3
Objects
2161
217100
Nodes
21710
7
Background--Small world paradigm--

• Small world networks represent two major
  properties
• Two randomly chosen nodes are connected by
  short avg. distance
• Nodes are joined together in groups. The effect
  is called high clustering.

8
Main Ideas

• Apply Small worlds small average distance to
  structure P2P routing in order to improve the
  performance of object lookup
• Apply Small worlds high clustering coefficient
  to provide large traffic resolving solution

9
Small world Overlay Protocol (SWOP)-- Overview --
Small World Overlay Layer
DHTs network
10
Small world Overlay Protocol (SWOP)-- Overview --

• Terminologies
• Type of links
• Short links
• Long links
• Types of nodes
• Head nodes
• Inner nodes

31
0
1
30
2
29
3
28
4
27
5
26
6
25
7
24
23
8
22
9
21
10
20
11
19
12
18
13
17
14
16
15
11
Join Cluster Protocol

• Compute distance between predecessor and
  successor nodes
• Retrieve group size from nodes
• Select a group to join
• Update links information

G1 2 G2 2
D1 3 D2 1
12
Leave Cluster Protocol

• Contact short links neighbors for leaving
• If it is the Head node, hand over the short
  links neighbor and long links neighbors to next
  Head, and generate necessary new long links
  neighbors

P.S. There exists boundary case, like only one
node in a cluster. The solution is written
in the thesis.
13
Leave Cluster Protocol

• Node 9 leaves
• Node 9 announces a new head 10 to 11
• Node 10 gets the long links and short links
• Link from node 4 to node 9 fade out when this
  link being used by a lookup request.

31
0
1
30
2
29
3
28
4
27
5
26
6
25
7
24
23
8
22
9
21
10
20
11
19
12
18
13
17
14
16
15
14
Object Lookup Protocol

• Phase 1
• Query short link neighbors if there exists one
  which contains desired object.
• If result is positive, the object lookup request
  ends. Otherwise, phase 2 begins.
• Phase 2
• By using the head s long link neighbor,
  forward the object lookup request to another
  cluster.
• Phase 1 continues by that node receives the
  object lookup request.

15
Object Lookup Example (1)Node 0 requests object
29, managed by node 31
16
Object Lookup Example (2)Node 0 requests object
16, managed by node 17
17
Performance analysis mathematical

• Worst case average link traversals
• EX (1 log2(m/2)) (8ln(3m)/k)
• EX represents worst case expected number of
  link traversals
• m represents the number of clusters in the SWOP
  network
• k represents the number of long links
• The proof is conducted by randomized algorithm
  and it leaves in the thesis

18
Performance analysis simulation result

• Performance Metrics
• Probability density function of lookup hop count
• Simulation Setup
• We added nodes one by one and reform the topology
  according the construction protocol.
• Total about 1k-5k nodes were generated in the
  system.
• Each node generated data lookup request randomly
  certain times.

19
Performance analysis simulation result
20
Internet Flash Crowd

• Unpredictable huge amount of request for a
  popular object is generated towards the object
  owner
• This overwhelms the network and the CPU resources
  of the owner.
• e.g. CNN news server during 911

21
Types for Internet Flash Crowd

• Static Flash Crowd
• The popular object involved will remain unchanged
  after its first appearance, e.g. new movie.
• Dynamic Flash Crowd
• The popular object involved will change after its
  first appearance, e.g. news

22
Requirements for handling Internet Flash Crowd

• Loading Distributed
• Reduce the bottleneck by caching and replication
• Demand driven
• Replication scheme has to be demand driven,
  otherwise, it will be a flooding scheme.
• Dynamic compatibility
• Consider how to handle dynamic objects which are
  changed by original source

23
Protocols for handlingFlash Crowd

• Algorithm
• Each node periodically records the access rates
  of objects stored.
• If the access rate of an object is greater than
  certain threshold, the owner of this object
  contacts the head of its cluster.
• head spread this object to all its long link
  neighbors

24
Protocols for handlingFlash Crowd

• Each long link neighbor caches this object so
  that each neighbor can be acted as an image
  source of this object.
• Each long link neighbor keeps track of the rate
  of newly added object.
• Dynamic cases Refresh Message
• Aims for reminding neighbor nodes to get the
  latest updated object.

25
Static Flash Crowd Example
Assume, in node 4, an items hitting rate exceeds
threshold
26
Dynamic Flash Crowd Example
27
Experiment results

• Study the performance on handling flash crowd
  scenario
• Performance metrics
• Number of successful request
• In each interval, the nodes try to lookup the
  popular object. The metric counts the number of
  nodes can retrieve in this interval.
• Number of messages produced (traffic burden)
• Total number of messages in the system.
• Settings
• Static flash crowd with fixed rate ? req/s.
• Dynamic flash crowd with fixed rate ? req/s with
  simulation time 25, 50 and 75 for object version
  update
• Dynamic flash crowd varying ? with simulation
  time 25, 50 and 75 for object version update
• Static flash crowd with fixed rate ? req/s

28
Experiment results

• Evaluation procedure
• We added 2000 nodes in the system using SWOP to
  form small world.
• One object was randomly chosen and was acted as
  the popular object.
• Each node generated that popular objects lookup
  request traffic with rate ?.
• Each node has fixed ? service rate and a queue
  for handling the item request.

29
Experimental resultsStatic flash crowd
30
Experimental resultsDynamic flash crowd
31
Experiment resultsVariation of object request
rate
32
Experiment resultsOperation cost
33
Summary

• Build a protocol that applies Small world
  features on P2P network
• Improve the Object Lookup performance with the
  support of mathematical analysis.
• Propose an algorithm to handle massive traffics
  produced by popular and dynamic objects.

34
Thank you

• Questions Answers

35
Small world Overlay ProtocolStabilize algorithm

• Method 1
• Each node periodically probes their neighbors.
  Once a timeout event occurs (node failure) for a
  corresponding probing, the routing information is
  updated.
• Method 2
• Each node does not perform a probing until an
  item lookup event occurs. When the lookup event
  fails, meaning there exists a node failure, the
  routing information is updated according to the
  failed node.

36
Experiment resultsVariation on number of long
link neighbors k
37
Experiment resultsVariation of queue size
38
Mathematical Analysis

• Apply Markov Model
• Define each finite state as number of cached
  clusters
• Retrieve the transition probabilities
• Define Troop State
• Compute the troop state probability
• Calculate the expected time to troop state