Measurements of Peer-to-Peer Systems

1
Measurements ofPeer-to-Peer Systems

• Pradnya Karbhari
• Nov 25th, 2003
• CS 8803 Network Measurements Seminar

2
Introduction to Peer-to-Peer (P2P) systems

• End-systems (or peers), are capable of behaving
  as clients and servers of data, hence system is
  scalable and reliable
• Peers participation is voluntary, membership is
  dynamic, hence topology keeps changing
• Most popularly used for file sharing, hence
  peer-to-peer systems have become synonymous with
  peer-to-peer file sharing networks

3
Classification of P2P systems

• P2P computation (e.g. seti_at_home)
• P2P communication (instant messaging)
• P2P file-sharing networks
• Centralized (e.g. Napster)
• Decentralized
• Structured (e.g. Chord, CAN, Pastry, Tapestry)
• Unstructured (e.g. Gnutella, Kazaa, Freenet,
  eDonkey, eMule, Direct Connect, )

4
Popularity of unstructured decentralized P2P
networks

• Gnutella host count, maintained by Limewire
  (http//www.limewire.com)
• good scope for measurement studies because
• deployed and widely used
• use a lot of bandwidth during data transfer,
  hence a concern for network operators
• quite a few measurement studies have been done on
  these systems, some of which we will discuss in
  this seminar

5
Outline

• Characterization of users of P2P systems
• Saroiu, et.al., A Measurement Study of
  Peer-to-Peer File Sharing Systems, MMCN, 2002.
• Effect of P2P traffic on the underlying network
• Sen, et.al., Analyzing peer-to-peer traffic
  across large networks, IMW02
• Peer-to-Peer Topologies
• Ripeanu, et.al., Mapping the Gnutella Network
  Properties of Large-Scale Peer-to-Peer Systems
  and Implications for System Design, IEEE
  Internet Computing, 2002.
• Searching on the P2P network
• Sripanidkulchai, The popularity of Gnutella
  queries and its implications on scalability,
  2001
• Deciphering proprietary P2P systems (like Kazaa)
• Leibowitz, et.al., Deconstructing the Kazaa
  Network, WIAPP, 2003.

6
Gnutella protocol overview

• Connecting to the Gnutella network
• bootstrap using GWebCache system and locally
  cached hostlist
• Ping/Pong messages are exchanged with potential
  neighbors
• Searching on the network
• Query messages are flooded on the network
• QueryHit messages are received (back-propagated
  along Query path) from peers having the requested
  content
• Downloading the content
• peers download files directly from peers having
  the requested content

7
Characterization of Users of P2P systems

• S. Saroiu, P. Gummadi and S. Gribble, A
  Measurement Study of Peer-to-Peer File Sharing
  Systems, MMCN02.
• first paper to characterize p2p file sharing
  systems
• Goal To analyze the following user
  characteristics
• latency
• lifetime of peers
• bottleneck bandwidth
• number of files shared and downloaded
• degree of cooperation
• methodology active crawling
• systems studied Napster and Gnutella
• data collection May 2001

8
Measurement Methodology

• active crawling of the Napster and Gnutella
  systems
• Napster issued queries for popular content, and
  then queried central server for peer information
• Gnutella used ping/pong messages in protocol to
  get metadata about peers, and then their
  neighbors and so on
• parallel measurement for
• peer lifetime- periodic probing of peers obtained
  from crawlers
• offline if no response to TCP SYN
• inactive if response to TCP SYN is a TCP RST
• active if accepts the incoming TCP connection on
  that port
• latency- RTT measurements from one host
• bottleneck link bandwidth- active probing using
  Sprobe, a tool they developed based on
  packet-pair dispersion technique

9
Host Lifetime analysis

• 20 peers in Napster, Gnutella have IP-level
  uptime of 93 or more
• Napster peers have higher application uptimes
  than Gnutella peers
• the best 20 of Napster peers have uptime of 83
  or more and the best 20 of Gnutella peers have
  uptime of 45 or more
• median session duration is 60 minutes for Napster
  and Gnutella

10
Latency analysis (Gnutella)

• 20 peers have a latency of at most 70ms and 20
  have a latency of at least 280ms
• correlation between downstream bottleneck
  bandwidth and latency two clusters for modems
  (20-60Kbps, 100-1000ms) and broadband (1Mbps,
  60-300ms)

11
Bottleneck Bandwidth Analysis (Gnutella)

• 92 Gnutella peers have downstream bottleneck
  bandwidth of at least 100Kbps
• 22 peers have upstream bottleneck bandwidth of
  100Kbps or less
• peers are unsuitable to serve content

12
Downloads, Uploads and Shared Files

• relative number of downloads and uploads varies
  significantly across bandwidth classes
• clear client/server behavior of different classes

13
Shared files v/s Shared Data(Napster and
Gnutella)

• Strong correlation between number of files shared
  and amount of shared MB of data
• slope of both lines is 3.7MB, the size of a
  typical MP3 audio file

14
Degree of Cooperation (Napster)

• 30 of the peers report bandwidth as 64Kbps or
  less, but actually have significantly higher
  bandwidths
• 10 of the peers reporting higher bandwidths
  (3Mbps or higher) actually have significantly
  lower bandwidth

15
Effect of P2P traffic on underlying network

• S. Sen and J. Wang, Analyzing peer-to-peer
  traffic across large networks, IMW 2002.
• Goal To characterize p2p traffic at three
  aggregation levels- IP, prefix and AS
• host distribution and host connectivity
• traffic volume and mean bandwidth usage
• traffic patterns over time
• connection duration and on-time methodology
  passive measurements at routers (port based)
• systems studied FastTrack(Kazaa), Gnutella,
  Direct Connect
• analysis of flow-level data collected from
  multiple border routers across a large tier-1
  ISPs backbone

16
Measurement Methodology

• flow records from multiple border routers
  matching ports
• 6346/6347 Kazaa
• 1214 FastTrack
• 411/412 Direct Connect
• processed data to eliminate
• private IP addresses
• invalid AS numbers
• final data set contained 800 million flow records

17
Datasets used for analysis

• FastTrack is most popular in terms of number of
  hosts participating and average traffic volume
  per day
• rapid growth of P2P traffic is mainly caused by
  increasing number of hosts in the system
• Direct Connect systems have higher traffic volume
  per IP address

18
Host distribution analysis

• of IP addresses in FastTrack ranges from 0.5 to
  2 million
• ratio of of IP addresses in FastTrackGnutellaD
  irectConnect is 150301
• Density of a prefix is the number of unique
  active IP addresses belonging to it
• Density of an AS is the number of unique prefixes
  belonging to it
• FastTrack hosts are distributed more densely than
  Gnutella and Direct Connect hosts (64164)

19
Host connectivity analysis (FastTrack)

• 48 of individual IPs communicate with at most
  one IP and 89 with at most 10 IPs
• 75 of prefixes and ASes communicate with at
  least 2 prefixes or ASes
• very few hosts have very high connectivity and
  most hosts have very low connectivity

20
Traffic volume analysis

• CDF of traffic volume per IP/prefix/AS for
  FastTrack (one day)
• distribution of P2P upstream traffic volume
  across three months

21
Mean bandwidth usage(FastTrack and Direct
Connect)

• FastTrack 33 IP addresses have mean downstream
  b/w 56Kbps or less 50 have mean upstream b/w
  56Kbps or less
• Direct Connect 20 IP addresses have mean
  downstream b/w 56Kbps or less 33 have mean
  upstream b/w 56Kbps or less

22
Traffic patterns over time (FastTrack)

• traffic volume transferred every hour among
  FastTrack hosts
• number of unique IP addresses, prefixes, ASes
  active every hour
• number of active unique IP addresses in each bin
  of various sizes
• system is very dynamic- hosts join and leave
  frequently

23
Connection duration and On-time (FastTrack)

• 50 of the IPs are online for less than one
  minute/day
• 60 IPs, 40 prefixes, 30 ASes stay for less
  than 10 mins/day
• 65 of the IPs join only once
• AS, prefix level- not very transient

24
Peer-to-Peer Topologies

• M. Ripeanu, I. Foster and A. Iamnitchi, Mapping
  the Gnutella Network Properties of Large-Scale
  Peer-to-Peer Systems and Implications for System
  Design, IEEE Internet Computing Journal, 2002.
• Goal To discover and analyze the Gnutella
  overlay topology and evaluate generated traffic
• methodology active crawling
• datasets Nov 2000, March 2001 and May 2001

25
Gnutella Network Growth

• number of nodes in the largest connected
  component in the Gnutella network
• significantly larger network found during
  Memorial Day and Thanksgiving
• 50 times increase within 6 months

26
Distribution of node-to-node shortest paths

• more than 95 node pairs are at most 7 hops away
• longest node-to-node path is 12 hops

27
Averag node connectivity

• average number of connections per node remains
  constant 3.4

28
Node connectivity distribution

• Nov 2000 Gnutella nodes organize themselves in a
  power law
• March 2001 connectivity does not look like a
  power law for all nodes power law distribution
  is preserved for nodes with more than 10 links
  for less than 10 links, the distribution is
  almost constant

29
Searching on the P2P network

• K. Sripanidkulchai, The popularity of Gnutella
  queries and its implications on scalability,
  2001, http//www-2.cs.cmu.edu/kunwadee/research/p
  2p/gnutella.html
• methodology passive measurements at one or two
  peers, made part of the Gnutella network, to log
  queries and query messages routed through it
• data sets Dec 2000, Jan 2001

30
Top 20 most popular query types

• 17 queries contained non-ASCII strings- filtered
  them out
• most queries for artists, adult content and file
  extensions (audio)
• some queries for books, software etc.

31
Query popularity distribution

• two distinct distributions of document
  popularity, with a break at query rank 100
• most popular documents are equally popular
• less popular documents follow a Zipf-like
  distribution, with alpha beween 0.63 and 1.24

32
Deciphering proprietary P2P systems

• Leibowitz, M. Ripeanu and A. Wierzbicki,
  Deconstructing the Kazaa Network, WIAPP, 2003.
• methodology passive content-based data
  collection at a caching server installed at the
  border of a large ISP
• L4 switch inspects first few packets of each TCP
  connection to detect Kazaa download traffic
• redirects Kazaa download traffic through caching
  server
• focus on download traffic only, not control
  traffic (since it is encrypted)

33
Characteristics of Collected Traces

• 38 of all download sessions do not use standard
  Kazaa port (1214)

34
File download distribution by bytes

• CDF of byte popularity distribution for 10, 1
  most popular files
• 0.8 of all files account for 80 of the
  generated traffic
• 0.1 of the most bandwidth hungry files (top 1
  of all files) generate 50 traffic

35
File size distribution

• note the log-scale on X-axis
• 3 distinct modes
• 100KB for pictures
• 2-5MB for music files
• 700MB for movies

36
Quantity and Rate of Distinct Files

• new files seen at different time scales- every
  day, hour, minute
• 150,000 distinct files during a 17-day period
• daily graph new files seen continued to
  decrease, but no steady state value (rate of
  injection of files in the network) achieved
• hourly graph time of day effect
• per-minute graph 50 new files seen every minute
  on an average

37
Rate of change of popularity of files

• percentage of files that make it to the N most
  popular files list- (a) in consecutive intervals
  and (b) after T intervals, compared with first
  list
• measurement interval is 24 hours
• 15 of the highly popular files remain popular
  throughout the experiment, and the rest are
  popular at short time intervals

38
Open Questions

• Mapping a global snapshot of the entire Gnutella
  topology
• Bootstrapping of peers in unstructured
  peer-to-peer systems (work in progress)
• More efficient searching on P2P networks- efforts
  in this direction include random walks,
  bloom-filter based techniques etc.
• End-point privacy/anonymity is absent in most of
  these peer-to-peer networks

39
References

• Papers covered in the seminar
• S. Saroiu, P. Gummadi and S. Gribble, A
  Measurement Study of Peer-to-Peer File Sharing
  Systems, MMCN 2002.
• S. Sen and J. Wang, Analyzing peer-to-peer
  traffic across large networks, IMW 2002.
• M. Ripeanu, I. Foster, A. Iamnitchi, Mapping the
  Gnutella Network Properties of Large-Scale
  Peer-to-Peer Systems and Implications for System
  Design, IEEE Internet Computing, 2002.
• Sripanidkulchai, The popularity of Gnutella
  queries and its implications on scalability,
  2001.
• N. Leibowitz, M. Ripeanu, A. Wierzbicki,
  Deconstructing the Kazaa Network, WIAPP 2003.
• Papers not covered in the seminar
• J. Chu, K.Labonte and B. Levine, Availability
  and Locality Measurements of Peer-to-Peer File
  Systems, SPIE, July 2002.
• F. Bustamante and Y. Qiao, Friendships that
  last Peer lifespan and its role in P2P
  protocols, WCW 2003.
• R. Bhagwan, S. Savage and G. Voelker,
  Understanding Availability, IPTPS 2003.
• Saroiu, et.al., An Analysis of Internet Content
  Delivery Systems, OSDI 2002.
• Markatos et.al., Tracing a large-scale
  Peer-to-Peer System An hour in the life of
  Gnutella, CCGrid 2002.