Measurement, Modeling, and Analysis of a PeertoPeer FileSharing Workload

1
Measurement, Modeling, and Analysis of a
Peer-to-Peer File-Sharing Workload

• K.P. Gummadi, R. J. Dunn, et al
• SOSP03
• Presented by Lu-chuan Kung kung_at_uiuc.edu

2
Outline

• Trace methodology and analysis
• User characteristics
• Client activities
• Object dynamics
• Analyze why Kazaa workload is not Zipf
• A model of P2P file-sharing workloads
• A study of bandwidth-saving techniques
• Conclusion

3
Trace Methodology

• Passively collect Kazaa traffic at the border of
  campus network and internet
• Query traffic was not captured b/c of encryption.
  File transfers are HTTP transfers w/
  Kazaa-specific header
• Summary statistics of the trace

4
Kazaa Users Are Patient

• Transfer time the difference between the start
  time and the end time of a request
• Small objects
• Large objects 100MB (typically video files)

5
User Slow Down As They Age

• Do people become hungrier for content as they
  gain experience with Kazaa?
• Older clients requested fewer bytes b/c
• Attrition population declines as clients age
• Slowing down older clients ask for less

6
Client Activity

• Its difficult to quantify the availability of
  clients in a p2p system
• Client activity includes
• Activity fraction time spent in transfers /
  duration of lifetime. Lower bound on availability
• Average session length typical duration length

7
Object Characteristics

• Kazaa is not one workload
• Kazaa is a blend of workloads of different
  properties
• 3 ranges of objects small ((10MB100GB), and large (100GB)
• Majority of requests are for smaller objects
• Most bytes transferred are due to large objects

8
Kazaa Object Dynamics

• Multimedia objects are immutable, therefore
  affect object dynamics
• Kazaa clients fetch objects at most once
• Kazaa client requests an object once 94 of time
• Kazaa client requests an object twice 99 of
  time
• Most requests are for old (repeated) objects
• An object is old if at least one month has passed
  since the first request of the object
• 72 of requests for large objects are old
• 52 of requests for small objects are old

9
Kazaa Object Dynamics

• The popularity of Kazaa objects is often
  short-lived
• The most popular pages remains stable for the Web
• Popularity is fleeting in Kazaa
• Audio files lose popularity faster than popular
  video files
• The most popular Kazaa objects tend to be
  recently born objects
• Newly born objects did not receive any requests
  during the first month of the trace

10
Kazaa Is Not Zipf

• Zipfs law
• The popularity of ith-most popular object is
  proportional to i-a, a Zipf coefficient
• Kazaa is not Zipf
• Most popular objects are less popular than Zipf
  would predict

11
Why Kazaa Is Not Zipf

• Fetch-repeatly vs. fetch-at-most-once
• Simulate the two cases based on the same Zipf
  distribution
• The result of fetch-at-most-once is similar to
  Kazaa.
• Non-Zipf workloads are also observed in web proxy
  caches and VoD servers

12
A Model of P2P File-Sharing Workloads

• Hypothesis underlying popularity of objects in a
  fetch-at-most-once system is driven by Zipfs law
• A client requests 2 objects per day. Choose which
  object to fetch from Zipf(1)
• An object is born with rate ?o , its popularity
  rank is selected from Zipf(1)
• Total object population cannot be observed from
  the trace. Use back-inference given 18,000
  distinct objects are requested in the trace,
  whats the total number of objects? Ans 40,000

13
Model Structure and Notation

• Parameter value are chosen to reflect the
  measured data from the trace

14
File-Sharing Effectiveness

• How should organization exepect bandwidth demand
  to change over time, given a shared proxy server?
• Hit rate of the proxy cache decreases in the
  fetch-at-most-once case
• Fetch-at-most-once clients consume the most
  popular objects early

15
New Object Arrivals Improve Hit Rate

• Object updates in Web lower the hit rate
• New objects arrivals are beneficial in P2P system
• Arrivals of popular objects increase hit rate
• If no arrivals, clients are forced to choose from
  the remaining unpopular objects

16
New Clients Cannot Stabilize Performance

• The infusion of new clients at a constant rate
  cannot compensate for the increasing number of
  old clients
• If we want to keep hit rate as a constant, we
  need exponential client arrival rate

17
Model Validation

• Underlying Zipf assumption cannot be validated
  directly.
• Use the proposed model to replicate the object
  popularity distribution in the trace
• Estimate various parameters
• Arrival rate of new objects is chosen to fit the
  measured data. ?o 5,475 objects per year

18
Exploring Locality-aware Request Routing

• A significant fraction of Internet bandwidth is
  consumed by Kazaa
• How would exploitation of locality help to save
  bandwidth?
• Different ways to exploit locality
• A centralized proxy cache placed at organization
  border
• Request redirection favor organization-internal
  peers
• Centralized request redirection
• Decentralized request redirection

19
An Ideal Proxy Cache

• Assume an ideal proxy infinite capacity and
  bandwidth
• 86 of external bandwidth would be saved
• However, some may not want to store P2P
  file-sharing content in a proxy server due to
  legal issues

20
Benefits of Locality-Awareness

• Trace-based simulation
• Infinite storage capacity
• At most 12 concurrent downloads
• Upload bandwidth 500 Kb/s
• External bandwidth 100 Kb/s
• Clients are available only when theyre
  transferring (a very conservative assumption)
• Cold misses objects cannot be found in peers
• Busy misses objects found but the peer is
  unavailable due to concurrent transfers

21
Benefits of Locality-Awareness

• Locality awareness obtained 68 byte hit rate for
  large objects and 37 byte hit rate for small
  objects
• A substantial number of miss bytes (62 of large
  objects, 43 of small objects) are due to
  unavailable clients

22
Benefits of Increased Availability

• Most of bytes served and consumed come from
  highly available peers
• Adding availability to the most available hosts
  earns a higher hit rate than adding to the least
  available host

23
Conclusion

• P2P file-sharing workloads are different to Web
  workloads
• User are patient
• Aged clients demand less
• Fetch-at-most once
• The proposed model suggests that client births
  and object births are the fundamental forces
  driving P2P workloads
• Theres significant locality in the Kazaa
  workload
• Locality-aware peers would save 63 external
  transfers even under conservation assumption

24
Comments

• Some of the observed characteristic may be
  related to the design of Kazaa and the measuring
  methodology and thus cannot be generalized
• The lack of portal sites in P2P system may also
  be a reason that most popular objects in P2P are
  less popular than Zipfs law would predict

25
Assessing the Quality of Voice Communications
Over Internet Backbones

• A.P. Markopoulou, F.A. Tobagi, M.J. Karam
• Tran. on Networking v11 no5 Oct 2003
• Presented by Lu-chuan Kung

26
Outline

• VoIP System
• Playout schemes
• Voice Impairment in Networks
• Internet measurements
• Numerical results
• Discussion

27
VoIP System
28
VoIP System

• Speech signal
• Talkspurts have mean 352ms
• Silence periods have mean 650ms
• Encoding schemes
• Packetizer add headers for different protocols
• Playout buffer packets are held for a later
  playout time in order to smooth playout
• Decoder reconstruct the speech signal

29
Playout Schemes

• Two types fixed and adaptive
• Fixed playout scheme
• End-to-end delay p is the same for all packets
• Large delay decreases packet loss due to late
  arrivals, but also decreases interactivity
• Adaptive playout scheme
• Estimate p based on delay dav and delay variation
  v
• p dav 4v
• Estimate p
• Talkspurt by talkspurt
• Packet by packet

30
Voice Impairment in Networks

• Quality of voice is affected by
• Encoding
• Packet loss
• Network delay jitter
• End-to-end delay
• Echo
• End-to-end delay consists of
• Encoding delay
• Packetization delay
• Network delay
• Playout bufferring delay
• Decoding delay

31
Assessment of Voice Communication in Packet
Networks

• Mean Opinion Score (MOS) a subjective rating
  given by listeners, given on a scale of 1-5
• Intrinsic quality MOSintr quality after
  compression

32
Degradation Due to Loss

• PLC Packet Loss Concealment
• Convert loss rate to MOS

33
Loss of Interactivity

• Loss of interactivity due to large end-to-end
  delay
• NTT study
• 6 conversation modes (tasks), task 1 is the
  hardest, task 6 is the most relaxed type

34
Echo Impairment

• Echo can cause major quality degradation
• The effect of echo is a function of delay and
  echo losses

35
Emodel

• Published by ITU-T. Provide formulas to predict
  MOS of voice quality
• R (R0 Is) Id Ie A
• R0 basic SNR
• Is impairment of signal, eg. sidetone and PCM
• Id impairment due to delay (echo
  interactivity)
• Ie impairment due to distortion (loss)
• A advantage factor (lenient users)

36
Internet Measurements

• Probe measurement
• 5 major U.S. cities
• 43 paths in total
• 7 providers P1,P2,,P7
• 50 bytes probes sent every 10 ms

37
Observations on the Traces

• Duration of the trace 3 days
• Network loss
• 6 out of 7 providers have outages
• Outages happened at least once per day
• Delay characteristics
• Delay spikes
• Alternation between high and low states
• Periodic clustered delay spikes

38
Delay Characteristics
39
Consistent Characteristics Per Provider
40
One Example Call

• Apply emodel to the traces using different
  playout buffer scheme
• Example of a 15-min call

41
One Example Call

• Fixed playout incurs many losses in the last 5
  mins

42
How to Choose p for Fixed Scheme

• Tradeoff between loss and delay
• There is a optimal value of delay to achieve
  maximum MOS value

43
Example Path Many Calls

• Random calls uniformly spread over an hour
• 150 short (3.5-min) and 50 long (10-min) calls
• Plot CDF vs. MOS

Fixed Playout
Adaptive Playout
44
Discussion

• Backbone networks have a wide range of
  performance
• Some are already able to support high quality
  voice communications
• Some are barely able to provide acceptable VoIP
  service (MOS 3.6)
• Reliability problems are more serious than QoS
  service mechanisms

45
Comments

• How representative are the chosen paths among the
  typical paths on Internet?
• End host to end host paths have larger delay