A Step Back Reflections on P2P

1
A Step BackReflections on P2P

• Boris Capitanu
• Ellick Chan
• 10/12/2004

2
2 P2P or Not 2 P2P?

• Mema Roussopoulos
• Mary Baker
• David S. H. Rosenthal
• TJ Giuli
• Petros Maniatis
• Jeff Mogul

3
Ideal P2P properties

• Self Organizing
• P2P routing
• Discovery
• Symmetric communication
• Peers are approximately equal
• Decentralized control
• No single point of failure

4
Constraints

• Budget
• Resource relevance to participants
• Trust
• Rate of system change
• Criticality
• Accountability
• Fault tolerance

5
Candidate problems

• Routing Problems
• Internet Routing (RON)
• Ad hoc Routing in Disaster Recovery
• Metropolitan-area Cell Phone Forwarding
• Backup
• Internet Backup
• Corporate Backup
• Distributed Monitoring

6
Candidate problems

• Data Sharing
• File sharing
• Censorship Resistance
• Data Dissemination
• Usenet
• Non-critical Content Distribution
• Critical Flash Crowds
• Auditing
• Digital Preservation
• Distributed Time Stamping

7
P2P models
Usenet
Gnutella
Images from http//www.cybergeography.org/atlas/mo
re_topology.html
8
2 P2P or not P2P
Budget
Relevance
Trust
9
Budget
Low
Effect
High

• Lowest possible cost per peer, rather than lowest
  global cost
• Bit Torrent, Gnutella, Freenet, etc.
• SETI_at_home

• Dictates how many peers join
• Decides if P2P is viable for problem

• Worries less about performance criticality
• Favors centralized approaches, P2P irrelevant
• Clusters, High performance computing

10
Relevance
Low
Effect
High

• Personal data
• Private data
• Internet backup
• Corporate backup
• Web caching

• Relevance of resources encourages peers to join
• When resource relevance is high, cooperation in
  a P2P solution evolves naturally

• File sharing
• Freenet
• Content distribution
• Internet routing
• Bit Torrent
• Gnutella
• Kazaa

11
Trust
Low
Effect
High

• Encryption
• Anonymity
• Freenet
• Oceanstore
• Ivy
• Timestamping
• MojoNation

• Mutual trust
• Risks

• Gnutella
• Napster
• Overlays
• File sharing
• Usenet

12
Rate of Change
Low
Effect
High

• Tangler
• Freenet
• LOCKSS
• Time stamping
• Content distribution
• Usenet
• Flash crowds

• Churn
• Timeliness
• Consistency

• Internet routing
• Online net monitoring

13
Criticality
Low
Effect
High

• Usenet
• Content distribution
• Offline net study
• File sharing

• Centralized control
• Accountability
• Fault tolerance

• Ad hoc disaster recovery
• Flash crowds
• Internet monitoring
• Routing

14
2 P2P or not P2P
Budget
Relevance
Trust
15
Conclusion

• Framework for analyzing P2P applications
• Captures constraints and app requirements
• Limited budget is motivating factor
• Problems with low relevance are inappropriate for
  P2P

16
Critique

• Strengths
• Quantifies application requirements and suitable
  use cases
• Generically describes suitability of classes of
  P2P apps
• Weaknesses
• Incomplete view of requirements
• Fuzzy requirements not accounted for

17
Service Capacity of Peer to Peer Networks

• Xianying Yang
• Gustavo de Veciana

18
Service Capacity

• of peers available to serve a document
• Throughput of P2P system
• Average delay
• Rate of dissemination
• What factors govern the effectiveness of a system
  to scale?

19
Research problem

• Analyze behavior of P2P systems
• Describe and model capacity behavior
• Transient regime
• Steady state
• Analyze conditions
• Does system scale as modeled?
• Are delays and throughput bounded?

20
Throughput

• Transient
• Steady state

21
Throughput
22
Service capacity model

• Steady state
• Impact of peer join/departure
• Performance
• Factors
• Peer selection
• Data management
• Multipart downloads
• Size of parts
• Admission and scheduling
• Traffic

23
Analysis

• 2 States
• Transient (branching process model)
• Steady state
• Deterministic
• Branching process
• Markov chain

24
Deterministic
0
Time
Rate
1

• N-1 users want a doc
• N2k
• S bits per request
• S(n-1) bits total
• Time interval ? at s/b seconds
• Exponential growth
• Ability to serve large bursts
• Average delays scales by lg(n)

0
Count
1
25
Deterministic
1
Time
Rate
1

• N-1 users want a doc
• N2k
• S bits per request
• S(n-1) bits total
• Time interval ? at s/b seconds
• Exponential growth
• Ability to serve large bursts
• Average delays scales by lg(n)

0
Count
2
1
26
Deterministic
2
Time
Rate
2

• N-1 users want a doc
• N2k
• S bits per request
• S(n-1) bits total
• Time interval ? at s/b seconds
• Exponential growth
• Ability to serve large bursts
• Average delays scales by lg(n)

0
Count
6
1
2
2
27
Deterministic
3
Time
Rate
4

• N-1 users want a doc
• N2k
• S bits per request
• S(n-1) bits total
• Time interval ? at s/b seconds
• Exponential growth
• Ability to serve large bursts
• Average delays scales by lg(n)

0
Count
8
1
2
3
3
2
3
3
28
Multipart

• M identical size chunks
• Service completions at s/mb??m seconds
• Optimization, peers favor others with no chunks
• At time k, system is partitioned into k sets
  Ai,i1k.
• Ai2k-i
• Ai corresponds to peers who have only received
  the ith chunk

A4
A2
A3
A1
Time slot k
29
Multipart

• M identical size chunks
• Service completions at s/mb??m seconds
• Optimization, peers favor others with no chunks
• At time k, system is partitioned into k sets
  Ai,i1k.
• Ai2k-i
• Ai corresponds to peers who have only received
  the ith chunk

A4
A2
A3
Time slot k
30
Peer groups
A1
A2
A3
A4
S
t0
t1
t2

tk
31
Multipart

• Delay is in effect reduced by a factor of m
• Large values of m better, but require more
  network overhead
• Congestion, bandwidth bottleneck ignored in this
  model

32
Branching Process Model

• Let Nd(t)peers serving document d at time t.
• Ti is a random variable, transfer time
• ET?1/?
• Age dependent branching process model, v2

33
Branching Process Model
34
Branching Process Model

• ????are growth characteristics
• If T is exponentially distributed, ??????????
• If T is deterministic, ?????ln2???????
• Exponential distribution increases growth exponent

35
Effect of v on Growth
Theorem II

• ??is inversely proportional to v
• Large fanout decreases growth exponent
• Intuition limit number of downloads at each peer

36
Peer Churn

• Peers exit system with probability 1-??upon
  completion
• If v??
• When peers exit, allowing multiple upload ensures
  document availability and system growth

System increases slowly with increasing v
37
Effect of m

• Allowing multipart downloads increases
  performance by factor m
• Growth rate increased by factor m
• Delay factor is reduced by 1/m
• Assumes peers are not simultaneously sharing
  multiple parts of files

38
Summary
Multipart
Branching
Deterministic

• Time interval ??for transfer
• N2k
• Delays bounded by ??log n
• Exponential growth

• Time interval ???m
• Delays bounded by (??m)?log n
• Space partitioned into sets
• More chunks is faster
• Network overhead is high

• Time interval ? is a random variable
• Delays bounded by log ?
• Parameters ??? determine operation
• Accounts for congestion, churn

39
Markov Chain

• Distant past irrelevant with knowledge of recent
  past
• Sequence of random variables, X1Xn
• Transition matrix
• Eigenvectors determine stable state conditions

40
Markov Chain
Sunny
Rainy
P(RainySunny)
Sunny
Rainy
P(RainyRainy)
P(SunnyRainy)
P(SunnySunny)
Weather, day 0
Weather, day 1
Weather, day 2
Weather, day n
41
Markov Model

• Poisson process r
• State
• xof peers requesting
• y peers hosting
• Multipart files
• Partial peers contribute at rate
• Total rate

i
Q
S0
Si
(1)
Exponentially distributed
Full service rate
Exit rate
42
Markov Model
43
Markov Model
44
Performance
Seeds/downloaders

• Seeds/downloaders
• ? is upload ratio of downloader to seed
• System with high ? leverages capacity
• Marginal change of system performance low when
  offered load is high

45
Bit Torrent

• Multipart d/l
• Chunk size 1 mb
• Credit system
• Updates every 5 min
• 150-200 file insertions

Service capacity
Throughput
Delay
46
Total Throughput
47
Average Throughput
48
Offered load and Delay
49
Throughput
50
Conclusions

• Credit system, growth are diametric
• Offered load linearly scales with number of peers
• Large multi-part files spread better
• Peer churn reduces throughput to constant
• Delays decrease with offered load

51
Scooped, Again

• Jonathan Ledlie, Jeff Shneidman, Margo Seltzer,
  John Huth

52
Outline

• Introduction
• Grid Computing
• P2P Systems
• Fallacies preventing cooperation
• Shared and Disjoint Problems
• Conclusions

What they are, Goals, Manifestations
53
Introduction

• Background, Motivation
• Peer-to-Peer vs. Grid Computing
• Overlapping problem domain
• P2P focuses on research
• Grid is concerned with concrete, tangible
  solutions
• History, repeated the Web

54
Introduction cont.

• Current trends
• Divergent, parallel development
• Duplication of work
• Grid risk of non-optimal solutions
• Missing out on P2Ps strong achievements (search
  and storage scalability, decentralization,
  anonymity, denial of service prevention)
• Cooperation is the key

55
Grids

• What is the Grid?
• a type of parallel and distributed system that
  enables the sharing, selection, and aggregation
  of resources distributed across multiple
  administrative domains based on the resources
  availability, capability, performance, cost, and
  users QoS requirements
• Short version virtualizing computer resources
• Large scale heterogeneous resource sharing
  (different platforms, hardware/software
  architectures, and computer languages)
• Functional classification
• Computational grids (run batch jobs during idle
  times)
• Data grids

56
Grid Layout
57
Grid Goals

• Design goal
• Solve problems too big for a single
  supercomputer, but retain the flexibility to work
  on multiple smaller problems
• Self-configuring, self-tuning, self-healing
• Allow data sharing and support computation across
  administrative domains
• Standardized programming interface
• GGF (Global Grid Forum)
• Globus toolkit the de facto standard for grid
  middleware

58
Grid Manifestations

• Protocols
• Resource management
• Grid Resource Allocation Management Protocol
  (GRAM)
• Information services
• Monitoring and Discovery Service (MDS)
• Security services
• Grid Security Infrastructure (GSI)
• Data movement and management
• Global Access to Secondary Storage (GASS),
  GridFTP
• Tools
• Grid Portal Software (GridPort, OGCE)
• Grid Packaging Toolkit
• Grid-enabled MPI (MPICH-G2)
• Network Weather Service
• Condor (CPU cycle scavenging) and Condor-G (job
  submission)
• APIs
• Web Services Open Grid Services Architecture
  (OGSA)

59
P2P

• What is P2P?
• a class of applications that take advantage of
  resources storage, cycles, content, human
  presence available at the edges of the
  Internet
• Decentralized, non-hierarchical node organization
• Inherently untrusted

60
P2P Goals

• Cost sharing / reduction
• Every peer responsible of its own cost
• Reduction of file storage costs
• Reduction of computation costs
• Improved scalability / reliability
• Lack of centralization allows new algorithms
  (CAN, Chordetc) to be designed to allow improved
  scalability
• Resource Aggregation
• Every peer lends its own resources to the network
• Increased Autonomy
• Tasks are performed locally no central service
  provider

61
P2P Goals cont.

• Anonymity / Privacy
• FreeNet
• Dynamism
• Nodes enter and leave the system in a transparent
  way
• Ad-hoc communication
• Members can join and leave based on their
  physical location or interests

62
Summary

• Grids
• Parallel, distributed systems concerned with
  resource sharing, selection, aggregation
• Resource availability, capability, performance,
  cost, and user QoS requirements are considered
• Self-configuring, self-tuning, self-healing
• Idle cycle and storage utilization

• P2P
• Distributed systems that take advantage of
  resources scattered throughout the Internet
• Decentralized, non-hierarchical node organization
• Concerned with fault-tolerance, scalability,
  availabilityetc.
• Idle cycle and storage utilization

63
Summary cont.

• Grid
• Distributed computation
• distributed.net
• SETI_at_home
• Data production / aggregation

• P2P
• Distributed file sharing
• Gnutella, KaZaA
• Distributed computation
• distributed.net
• Anonymity
• Freenet, Publius

64
Outline

• Introduction
• Grid Computing
• P2P Systems
• Fallacies preventing cooperation
• Shared and Disjoint Problems
• Conclusions

What they are, Goals, Manifestations
65
Fallacies preventing cooperation

• The technical problems in Grid systems are
  different from those in p2p systems
• Usage misconception Grid for computing problems,
  P2P for file sharing
• Data handling and data production in Grid systems
  has become important
• P2P used in desktop collaboration and network
  computation
• open problems in both camps have striking
  similarities

66
Fallacies preventing cooperation

• While the technical problems are similar, the
  architectures (physical topology, bandwidth
  availability and use, trust model, etc.) demand
  that the specific solutions be fundamentally
  different
• Solving common problems through sharing good
  ideas from each community
• Application dependent special requirements
  tailored to application needs, however the
  technical approaches for solving a particular
  problem could benefit both communities

67
Fallacies preventing cooperation

• Grid projects do not have the flexibility to try
  new algorithms/ideas because they have to get
  real work done. P2P research is all about this
  flexibility
• Grid has room for flexible research, too
• Testing new applications and protocols
• Users willing to adopt different technologies to
  get the work done

68
Outline

• Introduction
• Grid Computing
• P2P Systems
• Fallacies preventing cooperation
• Shared and Disjoint Problems
• Conclusions

What they are, Goals, Manifestations
69
Shared problems

• Topology Formation
• Node join and neighbor discovery
• Work has been done by both groups
• Grid On fully decentralized resource discovery
  in grid environments
• P2P Self-organization in p2p systems
• Grid infrastructure in not flexible hard coded
• Could benefit from P2P research prototypes

70
Shared problems cont.

• Utilization
• Resource discovery, data retrieval
• P2P hash-based look-up schemes are useful
• Resource management / optimization
• How to best utilize resources in a network
• Data replication/caching examined by both
  communities
• Scheduling and handling of contention
• P2P focus bandwidth usage (e.g. Gnutella)
• Grid focus scheduling
• Load balancing break large tasks into
  distributed smaller ones

71
Shared problems cont.

• Coping with Failure
• P2P lossy storage model (Freenet, Gnutella)
• Considerations for Grid adaptability
• Different common loss model
• Storage size (order of half a petabyte/month)
• Security-related issues
• Authenticity verification of data/computation
• Availability resilience to DoS attacks
• Authorization ACLs

72
Shared problems cont.

• Maintenance
• P2P essentially no standards or APIs
• Efforts by Berkeley BOINC, Google Compute,
  overlay standardization
• Grid pushes for a standardized API
• GGF (Global Grid Forum)
• OGSA (Open Grid Services Architecture)
• Web services oriented API Globus as reference
  implementation

73
Disjoint Problems

• Anonymity
• Not really useful for Grid systems, yet

74
Conclusions

• A lot of overlap between the goals and research
  interests of the two communities
• P2P community needs to consider the needs of the
  Grid users to see how existing research can be
  applied successfully to Grid problems
• Aim for common standards as much as possible

75
The Capacity of Wireless Networks

• Piyush Gupta, P. R. Kumar

76
Outline

• Introduction
• Arbitrary Networks
• Protocol and Physical Model
• Upper bound on transport capacity
• Constructive lower bound on transport capacity
• Random Networks
• Protocol and Physical Model
• Constructive lower bound on throughput capacity
• Possible Implications
• Discussion of tradeoffs
• Conclusion

77
Introduction

• Ad-hoc wireless networks
• No centralized control
• Each node involved in routing scheme
• Problems
• Network layer routing
• MAC varying network topology, decentralization
• TDMA too complex no centralized control
• FDMA inefficient in dense networks
• CDMA difficult to implement
• Random access preferred

78
Introduction cont.

• Sharing channels - hidden and exposed
  terminal problem MACA, MACAW use handshake
  signals to alleviate part of these problems
• Physical layer
• power regulated to minimize interference
• Exploring the capacity of wireless networks
• n nodes deployed in a 1 sq. meter region
• average distance between source and destination
  is L-bar
• Bandwidth each node can transmit at W bps over
  common wireless channel
• Multi-hop transmission with buffering
• Two types
• Arbitrary Networks
• Random Networks

79
Arbitrary Networks

• Nodes are arbitrarily distributed over a unit
  area disc
• Destination is arbitrary
• Rate is arbitrary
• Transmission range is arbitrary
• How can we model if a transmission was received
  successfully by the receiver?
• Two models Protocol Model, Physical Model

80
Protocol Model

• Transmission from Xi to Xj is successful if for
  every node Xk transmitting simultaneously
• where Xi denotes the location of a node, and ?
  is the guarding zone specified by the protocol

81
Physical Model

• - subset of nodes
  transmitting simultaneously at some time instant
  over a certain sub-channel
• Pk power level chosen at node Xk
• A transmission originating at node Xi is
  successfully received at node Xj if
• ß minimum signal-to-interference ratio
  necessary for successful reception
• N ambient noise power level
• a 2
• Signal power decays with distance as 1/ra

82
Transport Capacity of Arbitrary Networks

• Bit-meter a bit transported a distance of 1m
• used as indicator of a networks transport
  capacity
• Protocol Model
• Main result
• If this capacity is divided between the n nodes,
  we have
• for each node
• For equidistant destinations, the throughput
  capacity is

83

• Physical Model
• Main result
• is feasible
• is not
  feasible
• for appropriate values of c, c

84
Upper bound on transport capacity

• Assumptions
• There are n nodes arbitrarily located in a planar
  disk of unit area
• The network transports lnT bits over T seconds
  (each node generates bits at rate l bps)
• The average distance between source and
  destination of a bit is L
• Transmissions are slotted into synchronized slots
  of length t seconds

85
Upper bound on transport capacity

• Protocol Model
• Physical Model
• If Pmax/Pmin

86
Constructive lower bound on transport capacity

• Theorems and Lemmas that show a scenario where
  the order of the upper bound presented earlier is
  achieved
• There exists a placement of nodes and assignment
  of traffic patterns such that the network can
  achieve
• under Protocol Model, and
• under Physical Model
• Proofs are in the paper

87
Outline

• Introduction
• Arbitrary Networks
• Protocol and Physical Model
• Upper bound on transport capacity
• Constructive lower bound on transport capacity
• Random Networks
• Protocol and Physical Model
• Constructive lower bound on throughput capacity
• Possible Implications
• Discussion of tradeoffs
• Conclusion

88
Random Networks

• n nodes randomly located on the surface of a
  sphere of area 1 sq. meter (S2), or disk of area
  1 sq. meter in the plane
• independently and uniformly distributed
• randomly chosen destination with send rate l(n)
  bps
• assumptions all nodes are homogeneous (all
  transmissions employ the same nominal range or
  power)
• Two models
• Protocol model, Physical model

89
Protocol Model

• A transmission from Xi reaches Xj successfully if
  for every other Xk transmitting, the following
  holds
• 1.
• 2.
• where Xi represents the location of a node and r
  is the common range

90
Physical Model

• - subset of nodes transmitting
  simultaneously at some time instant over a
  certain sub-channel
• Let P be the common power level
• Then, a transmission from a node Xi is
  successfully received by node Xj if

91
Throughput Capacity of Random Networks

• Feasible throughput
• if a transmission schedule can be achieved such
  that every node can send l(n) bits/sec on average
  to its destination node
• depends on the location of nodes (random)
• Result
• Protocol model
• Physical model

92
Constructive lower bound on throughput capacity

• Goal show that virtual channel capacity
  guarantee of each source-destination pair of
  randomly located nodes is
• with probability approaching 1 as
  for c 0
• Steps
• define a Voronoi tessellation of S² where each
  cell is carefully chosen in relation to the
  number of nodes
• bound the number of interfering neighbors of a
  Voronoi cell
• bound the length of an all-cell transmission
  schedule
• define the routes of a packet in the Voronoi
  tessellation
• prove that each cell contains at least one node
• calculate the expected routes that pass through a
  cell and infer the expected traffic of each node

93
Outline

• Introduction
• Arbitrary Networks
• Protocol and Physical Model
• Upper bound on transport capacity
• Constructive lower bound on transport capacity
• Random Networks
• Protocol and Physical Model
• Constructive lower bound on throughput capacity
• Possible Implications
• Discussion of tradeoffs
• Conclusion

94
Possible Implications

• Results allow for a perfect scheduling algorithm
  that knows the location of all nodes and traffic
  demands, and coordinates the wireless
  transmissions temporally and spatially to avoid
  collisions (however, if the nodes are mobile or
  location information is not available, the
  capacity can only be smaller)
• As the number of nodes increases, the throughput
  decreases
• Feasible scenario If communication occurs only
  between nearby nodes, the bit rate does not
  decrease with n
• Scaled distance between source and destination is
  O(1/sqrt(n)) meters
• Power consumption
• Faster rate of decay of signal power with
  distance allows greater transport and throughput
  capacity

95
Implications cont.

• Division of labor is possible
• One node in a cell can be designated to relay
  multi-hop packets, if desired
• Tradeoffs upper bound on throughput
• Conflict between reducing the number of hops and
  increase spatial concurrency and frequency reuse
• Must reduce r(n) to smallest value possible
  without losing connectivity

96
Tradeoffs cont.

• Arbitrary Networks under the Protocol model
• Constraints that determine the transport capacity
  to be at most
• The length of routes
• Consumption of two-dimensional area by
  transmission
• Total number of nodes

97
Conclusions

• Designers may want to consider designing networks
  with small number of nodes
• Communication with nearby nodes at constant bit
  rates can be provided in a dense clusters of
  nodes, since the source destination distance
  shrink as O(1/sqrt(n))

98
Appendix A spatial tessellation

• Voronoi tessellation of the surface of the S²
  sphere
• A Voronoi cell is the set of all points which are
  closer to ai than to any of the other ajs
• Adjacent cells share a common point
• Every node in a cell is within distance r(n) of
  every node in own cell or adjacent cell
• Interfering neighbors a point in one cell is
  within a distance (2?)r(n) of some point in the
  other cell

99
Tessellation Properties

• For each e0, there is a Voronoi tessellation
  such that Each cell contains a disk of radius e
  and is contained in a disk of radius 2e
• Every Voronoi cell contains a disk of area
• with radius r(n)
• Every Voronoi cell is contained in a disk of
  radius 2r(n)

100
Bound on number of interfering neighbors of a cell

• Every cell in Vn has no more than c1 interfering
  neighbors
• c1 f(?) and grows linearly in (1?)²
• Allows construction of a schedule of bounded
  length
• Each cell in the tessellation Vn has an
  opportunity to transmit every 1c1 slots such
  that transmission is successful within a r(n)
  distance from the transmitter (in the Protocol
  Model)