IEG5270 Advanced Topics in P2P Networking Introduction

1
IEG5270 Advanced Topics in P2P
NetworkingIntroduction

• Dah Ming Chiu
• Chinese University of Hong Kong

2
What is P2P networking?

• What P2P applications are you aware?
• What distinguishes a P2P application from a
  non-P2P application?
• Why are we interested in P2P networking?

3
P2P traffic in Internet
4
Client-server applications

• Traditional applications are all client-server
• A service provider must set up a server e.g.
• Email server
• Web server
• A client is configured with servers address/port
• Server responds to clients one by one

Server
5
Limitations of client/server appls

• Scalability
• Compute power, access bandwidth, storage
• Costs
• Server, bandwidth, power, management
• Privacy concerns
• Some are for illegal reasons
• The need for fixed address/port

6
Peer-to-peer applications

• In a pure p2p network, every node is both a
  client and a server
• As clients increases, servers also increases
• perfectly scalable
• Also
• Distributes costs
• Increases privacy
• May use dynamic addresses

7
Hurdles for p2p applications

• How to find where things are?
• instead of static configuration of servers
• How to make peers able to help each other?
• A peer usually need to acquire the right content
  to be able to serve others
• How to deal with unreliable peers?
• Peers come and go Churn
• How to make peers willing to help each other?
• Incentives
• Why should other peers be trust-worthy?

8
A short history of some note-worthy p2p
applications

• Napster (1999-2002)
• First p2p file sharing application
• Relies on a centralized directory to help find
  content
• Once peer X knows peer Y has what it wants, X
  contacts Y directly
• X and Y often exchange copy-right protected
  material
• Napster server shutdown due to law suits
• Another company bought Napster, still use the
  name to operate on-line music shop

9
History - cont

• Gnutella (2000-?)
• Completely distributed p2p file sharing
• Peers form an overlay network each peer knows
  its neighbors
• Each peer floods its request to all other peers
• Prohibitive overheads
• Open source

10
History - cont

• Kazza or FastTrack (2001-2003)
• Partly centralized, rely on super-nodes
• Also met with law suits
• Skype (2003-now)
• Created by same folks who created Kazza
• Use similar technology as Kazza to discover VoIP
  destinations, instead of illegal content
• Plus other bells and whistles (codec, encryption,
  instant messaging, friends list/presence)
• Provide relay service if firewalls/NAT prevent
  connectivity

11
History - cont

• BitTorrent (2002-now)
• Does not deal with the content discovery problem
  (avoiding legal problem)
• Solves another problem how to distribute a file
  to a (large) number of peers
• Divide a file into many parts chunks
• Give chunks to different peers
• Peers rely on tracker to find other peers and
  what chunks they have
• Content flows through multiple, ad-hoc trees
• Several variations, and several other protocols
  (e.g. eDonkey)

12
History - cont

• PPLive (2005-now)
• Use similar technology as BT to stream Video (TV)
• Need to have a directory of TV channels/programs
• Some earlier start-ups had legal problems
• Mainly popular in China
• 10s of millions of steady customers
• Several competitors PPStream, UUC

13
Summary

• Success stories
• Skype p2p VoIP
• BitTorrent p2p file sharing
• PPLive p2p streaming
• Some other Client-server appls may change to p2p
  model
• Microsofts study of using P2P for VoD service
  (Sigcomm 2007)
• Google to use p2p for YouTube?

14
Other notable P2P projects

• LOCKSS (Lots of Copies Keep Stuff Safe)
• Data preservation mechanism (suppose libraries
  want to keep books preserved, even when
  publishers die)
• Data replicated many times
• How to deal with sabotage? Peers do periodic
  voting to validate all copies.

15
Other P2P projects

• SETI_at_home (search for Extra-Terrestrial
  Intelligence)
• Peers devote their compute power to help analyze
  radio signals from a telescope, to check for
  signs of ET
• P2P because peers are autonomous, operation is
  distributed
• A grid computer really

16
Other P2P projects

• RON (Resilient Overlay Network)
• In the Internet, packets do not always flow along
  the shortest paths due to ISP peering
• Peers help relay packets to support shortest path
  routing as much as possible
• Peers form an overlay network, and check delay
  between each pair of neighboring nodes
• RON demonstrated that it can provide lower delay
• Overlay versus p2p

17
Instant messaging vs p2p

• Some consider ICQ as the first p2p application
• In some sense, it is similar to Skype
• The communication part is p2p
• The search part may be partly via a centralized
  server
• Instant messaging may be integrated with file
  sharing and streaming, as well as VoIP (e.g. QQ
  and QQlive, a p2p application in China)

18
Summary of other p2p appls

• The idea of P2P overlaps with application layer
  infrastructures, e.g.
• Grid computing
• Overlay networks
• Distributed databases, search
• DHT is a building block for many such
  infrastructures

19
Important building blocks

• How to find things?
• Centralized approach
• Flooding
• Partially distributed approach
• DHT
• How to distribute content to multiple peers?
• IP multicast
• Structured tree(s) (or push)
• Unstructured (or data-driven, or pull)

20
Distributed Hash Tables

• Each object (e.g. file, chunk) is mapped to a key
• The key space is partitioned among the peers
• Given a key, you know where to store/retrieve the
  object
• There are quite a few DHT proposals
• Chord, CAN, pastry
• Every DHT supports one main operation
• Given a key, route to a peer that holds the key

21
DHT cont

• A DHT should have the following properties
• Decentralization no central management needed
• Scalability supports millions of nodes
• Reliability still works when peers leave and
  join
• Other properties
• Small diameter, small degree, get to destination
  node in log(n) steps
• Compared to index server no single point of
  failure
• Compared to flooding more efficient
• Limitation supports lookup, not search

22
The magic of streams

• Traditional multicast
• Single tree
• receivers do not contribute
• P2P content distribution
• use multiple trees to distribute chunks
• peers all contribute uplink bandwidth
• Peers are unreliable
• use data-driven approach to build distribution
  trees dynamically

23
Summary of building blocksto be covered in this
course

• DHT
• Multi-tree content distribution
• Models for understanding the capacity achieved by
  these algorithms
• Others incentive systems etc

24
The economics of p2p applications

• P2P applications induce a huge amount of traffic
  for ISPs
• P2P applications can often extract profit from
  users in one way or another
• Skype provide gateway into phone networks to
  collect money
• Streaming and VoD service can generate targeted
  advertising opportunities (the same game Google
  plays)

25
Net neutrality

• ISPs are often prevented from sharing the profits
  from content providers (who use p2p technology)
• To extract such profit, ISPs need almost monopoly
  power
• Governments do not want ISPs to become monopolies
• net neutrality is implicit business model
• Charge users based on volume of bits, not on
  content

26
The tussle between ISPs and P2P

• P2p file sharing or streaming is determining its
  own routes for content to travel
• ISPs settlements depend on their roles in peering
  relationships
• E.g. customer ISP pays both transit providers,
  but may be transiting some p2p traffic for them

Transit Provide1
Transit Provide2
Customer ISP
27
Tussle (conflict)

• How should ISPs consider peering decisions in
  view of p2p traffic?
• How should ISPs provision their peering
  bandwidth?
• How should ISPs charge subscribers (peers)?
• How should ISPs manage its p2p traffic?

28
P2P traffic monitoring and detection

• Traditional ISP settlements based on using
  netflow
• Gives total traffic volume, and volume for each 5
  minute interval
• ISP can see traffic types based on well-known
  ports
• P2p flows may not use well-known ports
• Deep packet inspection check signature in
  payloads
• What if payload is encrypted?

29
Detecting P2P traffic by flow behavior

• How many other nodes a node talks to
• What combination of protocols is used (UDP, TCP
  etc)
• Patters on packet sizes
• Active research area

30
ISP-friendly P2P

• ISPs love and hate p2p applications at the same
  time
• P2p applications are the killer appls for
  broadband access drawing and keeping customers
• P2p traffic quickly eats up any added bandwidth
• ISPs and researchers are looking for p2p
  algorithms and ISP caching services to reduce the
  cross-ISP traffic demands

31
Discussion the future of p2p

• Can it ever be made reliable enough as a service?
• May need to deploy smart servers to help p2p when
  necessary, e.g. in PPlive.
• Will ISPs change their traffic controls or
  charges to kill p2p?
• Unlikely in the near term ISP need content
  distribution applications
• P2p is quite effective, and can be made more
  ISP-friendly

32
Commercial interest in P2P

• Currently there is a lot of commercial interest
  in p2p technology
• Large investment in Joost
• Many other start-ups
• Interest by Microsoft and Google

33
Research interest in p2p

• Content distribution algorithms
• For file sharing
• For streaming
• For Video on Demand (VoD)
• DHTs and their use
• Revisit naming, addressing and routing in the
  next generation Internet
• Incentive and reputation systems
• No good incentive system for p2p streaming

34
Research interests - cont

• Applying network coding to p2p
• It may not help improve optimal throughput
• But it can help simplify peer and chuck selection
  strategies in distribution algorithms
• P2P traffic classification

35
Compare to a regular course

• There are some theoretical foundations, but less
  mature than a regular course
• The long term prospect may not be so clear, but
  the current interest is high
• Although undergraduates are allowed, but will run
  like a graduate course
• No exams, no text book
• Read papers, do some projects, learn to do
  research
• Practice oral presentation and writing

36
Homework projects

• Streaming algorithm design and simulation
• Based on YP Zhous research paper (ICNP 2007)
• Under some assumptions, try to design the best
  chunk selection strategy
• Simulate it and compare to YPs algorithm
• P2P traffic trace analysis
• Be a detective find p2p traffic in a trace
• Find properties of different p2p applications
• Planet Lab - ?

37
Individual project

• If too many students, we need to make them group
  projects (2 in a group)
• Three types I can see
• Survey a problem in depth (read several papers
  related to the problem and summarized/discuss)
• Do some research on a given problem
• Implement some specific P2P application/mechanism
  and demo it.
• Will give a list of potential topics
• Will give a list of papers to read

38
Some example topics

• Most commercial systems (e.g. PPlive) are not
  based on open source. Try to deduce the algorithm
  for specific p2p streaming applications
• What is P2P-SIP and what are its applications?
• Is it possible to implement a viable search
  engine based on p2p technology?

39
Assessment

• Main project 50
• Oral presentation written report
• Homework projects 40
• Class participation 10