Brief Overview of Academic Research on P2P

1
Brief Overview of Academic Research on P2P

• Pei Cao

2
Relevant Conferences

• IPTPS (International Workshop on Peer-to-Peer
  Systems)
• ICDCS (IEEE Conference on Distributed Computer
  Systems)
• NSDI (USENIX Symposium on Network System Design
  and Implementation)
• PODC (ACM Symposium on Principles of Distributed
  Computing)
• SIGCOMM

3
Areas of Research Focus

• Gnutella-Inspired
• The Directory Service Problem
• BitTorrent-Inspired
• The File Distribution Problem
• P2P Live Streaming
• P2P and Net Neutrality

4
Gnutella-Inspired Research Studies

5
The Applications and The Problems

• Napster, Gnutella, KaZaa/FastTrak, Skype
• Look for a particular content/object, and find
  which peer has it ? the directory service
  problem
• Challenge how to offer a scalable directory
  service in a fully decentralized fashion
• Arrange direct transfer from the peer ? the
  punch a hole in the firewall problem

6
Decentralized Directory Services

• Structured Networks
• DHT (Distributed Hash Tables)
• Very active research areas from 2001 to 2004
• Limitation lookup by keys only
• Multi-Attribute DHT
• Limited support for query-based lookup
• Unstructured Networks
• Various improvements to basic flooding based
  schemes

7
What Is a DHT?

• Single-node hash table
• key Hash(name)
• put(key, value)
• get(key) -gt value
• How do I do this across millions of hosts on the
  Internet?
• Distributed Hash Table

8
Distributed Hash Tables

• Chord
• CAN
• Pastry
• Tapastry
• Symphony
• Koodle
• etc.

9
The Problem
N2
N1
N3
Internet
Put (Keytitle Valuefile data)
?
Client
Publisher
Get(keytitle)
N6
N4
N5

• Key Placement
• Routing to find key

10
Key Placement

• Traditional hashing
• Nodes numbered from 1 to N
• Key is placed at node (hash(key) N)
• Why Traditional Hashing have problems

11
Consistent Hashing IDs

• Key identifier SHA-1(key)
• Node identifier SHA-1(IP address)
• SHA-1 distributes both uniformly
• How to map key IDs to node IDs?

12
Consistent Hashing Placement
A key is stored at its successor node with next
higher ID
13
Basic Lookup
14
Finger Table Allows log(N)-time Lookups
½
¼
1/8
1/16
1/32
1/64
1/128
N80
15
Finger i Points to Successor of n2i
N120
112
½
¼
1/8
1/16
1/32
1/64
1/128
N80
16
Lookups Take O(log(N)) Hops
N5
N10
N110
K19
N20
N99
N32
Lookup(K19)
N80
N60
17
Chord Lookup Algorithm Properties

• Interface lookup(key) ? IP address
• Efficient O(log N) messages per lookup
• N is the total number of servers
• Scalable O(log N) state per node
• Robust survives massive failures
• Simple to analyze

18
Related Studies on DHTs

• Many variations of DHTs
• Different ways to choose the fingers
• Ways to make it more robust
• Ways to make it more network efficient
• Studies of different DHTs
• What happens when peers leave aka churns?
• Applications built using DHTs
• Tracker-less BitTorrent
• Beehive --- a P2P based DNS system

19
Directory Lookups Unstructured Networks

• Example Gnutella
• Support more flexible queries
• Typically, precise name search is a small
  portion of all queries
• Simplicity
• High resilience against node failures
• Problems Scalability
• Flooding ? of messages O(NE)

20
Flooding-Based Searches
1
3
2
4
6
5
7
8
. . . . . . . . . . . .

• Duplication increases as TTL increases in
  flooding
• Worst case a node A is interrupted by N q
  degree(A) messages

21
Problems with Simple TTL-Based Flooding

• Hard to choose TTL
• For objects that are widely present in the
  network, small TTLs suffice
• For objects that are rare in the network, large
  TTLs are necessary
• Number of query messages grow exponentially as
  TTL grows

22
Idea 1 Adaptively Adjust TTL

• Expanding Ring
• Multiple floods start with TTL1 increment TTL
  by 2 each time until search succeeds
• Success varies by network topology

23
Idea 2 Random Walk

• Simple random walk
• takes too long to find anything!
• Multiple-walker random walk
• N agents after each walking T steps visits as
  many nodes as 1 agent walking NT steps
• When to terminate the search check back with the
  query originator once every C steps

24
Flexible Replication

• In unstructured systems, search success is
  essentially about coverage visiting enough nodes
  to probabilistically find the object gt
  replication density matters
• Limited node storage gt whats the optimal
  replication density distribution?
• In Gnutella, only nodes who query an object store
  it gt ri ? pi
• What if we have different replication strategies?

25
Optimal ri Distribution

• Goal minimize ?( pi/ ri ), where ? ri R
• Calculation
• introduce Lagrange multiplier ?, find ri and ?
  that minimize
• ?( pi/ ri ) ? (? ri - R)
• gt ? - pi/ ri2 0 for all i
• gt ri ? ? pi

26
Square-Root Distribution

• General principle to minimize ?( pi/ ri ) under
  constraint ? ri R, make ri proportional to
  square root of pi
• Other application examples
• Bandwidth allocation to minimize expected
  download times
• Server load balancing to minimize expected
  request latency

27
Achieving Square-Root Distribution

• Suggestions from some heuristics
• Store an object at a number of nodes that is
  proportional to the number of node visited in
  order to find the object
• Each node uses random replacement
• Two implementations
• Path replication store the object along the path
  of a successful walk
• Random replication store the object randomly
  among nodes visited by the agents

28
KaZaa

• Use Supernodes
• Regular Nodes Supernodes 100 1
• Simple way to scale the system by a factor of 100

29
BitTorrent-Inspired Research Studies

30
Modeling and Understanding BitTorrent

• Analysis based on modeling
• View it as a type of Gossip Algorithm
• Usually do not model the Tit-for-Tat aspects
• Assume perfectly connected networks
• Statistical modeling techniques
• Mostly published in PODC or SIGMETRICS
• Simulation Studies
• Different assumption of bottlenecks
• Varying details of the modeling of the data
  transfer
• Published in ICDCS and SIGCOMM

31
Studies on Effect of BitTorrent on ISPs

• Observation P2P contributes to cross-ISP traffic
• SIGCOMM 2006 publication on studies in Japan
  backbone traffic
• Attempts to improve network locality of
  BitTorrent-like applications
• ICDCS 2006 publicatoin
• Academic P2P file sharing systems
• Bullet, Julia, etc.

32
Techniques to Alleviate the Last Missing Piece
Problem

• Apply Network Coding to pieces exchanged between
  peers
• Pablo Rodriguez Rodriguez, Microsoft Research
  (recently moved to Telefonica Research)
• Use a different piece-replication strategy
• Dahlia Makhi, Microsoft Research
• On Collaborative Content Distribution Using
  Multi-Message Gossip
• Associate age with file segments

33
Network Coding

• Main Feature
• Allowing intermediate nodes to encode packets
• Making optimal use of the available network
  resources
• Similar Technique Erasure Codes
• Reconstructing the original content of size n
  from roughly a subset of any n symbols from a
  large universe of encoded symbols

34
Network Coding in P2P The Model

• Server
• Dividing the file into k blocks
• Uploading blocks at random to different clients
• Clients (Users)
• Collaborating with each other to assemble the
  blocks and reconstruct the original file
• Exchanging information and data with only a small
  subset of others (neighbors)
• Symmetric neighborhood and links

35
Network Coding in P2P

• Assume a node with blocks B1, B2, , Bk
• Pick random numbers C1, C2, , Ck
• Construct new block
• E C1 B1 C2 B2 Ck Bk
• Send E and (C1, C2, , Ck) to neighbor
• Decoding collect enough linearly independent
  Es, solve the linear system
• If all nodes pick vector C randomly, chances are
  high that after receiving K blocks, can recover
  B1 through Bk

36
P2P Live Streaming

37
Motivations

• Internet Applications
• PPLive, PPStream, etc.
• Challenge QoS Issues
• Raw bandwidth constraints
• Example PPLive utilizes the significant
  bandwidth disparity between Univeristy nodes
  and Residential nodes
• Satisfying demand of content publishers

38
P2P Live Streaming Cant Stand on Its Own

• P2P as a complement to IP-Multicast
• Used where IP-Multicast isnt enabled
• P2P as a way to reduce server load
• By sourcing parts of streams from peers, server
  load might be reduced by 10
• P2P as a way to reduce backbone bandwidth
  requirements
• When core network bandwidth isnt sufficient

39
P2P and Net-Neutrality

40
Its All TCPs Fault

• TCP per-flow fairness
• Browsers
• 2-4 TCP flows per web server
• Contact a few web servers at a time
• Short flows
• P2P applications
• Much higher number of TCP connections
• Many more endpoints
• Long flows

41
When and How to Apply Traffic Shaping

• Current practice application recognition
• Needs
• An application ignostic way to trigger the
  traffic shaping
• A clear statement to users on what happens