Measurement, Modeling, and Analysis of a Peer-2-Peer File-Sharing Workload

1
Measurement, Modeling, and Analysis of a
Peer-2-Peer File-Sharing Workload

• Presented For
• Cs294-4 Fall 2003
• By Jon Hess

2
Measurement, Modeling, and Analysis
of a Peer-2-Peer File-Sharing Workload

• Goal - Overview
• Determine if the KaZaA search space is queried in
  such a way that a group of 25,000 clients can
  satisfy most of their own requests.

3
Measurement, Modeling, and Analysis
of a Peer-2-Peer File-Sharing Workload

• Goals - Details
• Capture an extensive trace
• Utilize that trace to understand file-sharing
  traffic flows
• Model user and object activity
• Determine inefficiencies in the distribution
  model
• Propose solutions to inefficiencies

4
Measurement, Modeling, and Analysis
of a Peer-2-Peer File-Sharing Workload

• Motivations
• Beginning in 1999-2000 file-sharing traffic began
  to exceed HTTP traffic in terms of aggregate
  bandwidth consumed
• File-sharing traffic is much less understood than
  HTTP traffic even though it represents such a
  large segment of bandwidth usage
• Bandwidth is expensive

2000
2002
HTTP Traffic
P2P Traffic
5
Measurement, Modeling, and Analysis
of a Peer-2-Peer File-Sharing Workload

• The Trace
• 2 Machines
• 203 days 5 hours and 6 minutes
• 22.7TB of KaZaA file transfer traffic
• Captured seasonal variations
• End of spring
• Summer
• Fall semester

6
Measurement, Modeling, and Analysis
of a Peer-2-Peer File-Sharing Workload

• Trace Conclusions
• Users are patient
• 30 minutes to retrieve a small object
• Up to 1 week to retrieve a large object
• Users consume less as they age

7
Measurement, Modeling, and Analysis
of a Peer-2-Peer File-Sharing Workload

• Trace Conclusions
• Users machines are not very active
• A session is an unbroken length of time where a
  client has one or more file transfers in
  progress.
• Average sessions are only 2 minutes
• 90th percentile 28 minutes
• Over the life of a client, it is only active
  5.54 of the time or 0.20 of the trace period
• 90th percentile clients are active most of their
  life, and 4.15 of the trace
• Without control traffic analysis, is this
  meaningful?

8
Measurement, Modeling, and Analysis
of a Peer-2-Peer File-Sharing Workload
Session Lengths
9
Measurement, Modeling, and Analysis
of a Peer-2-Peer File-Sharing Workload

• Trace Conclusions Objects
• Most requests are for small objects 91
• Most bytes transferred are part of large objects
  65
• There are many small objects
• There are few large objects
• Small Objects popularity is subject to heavy
  churn
• No small object was in the top 10 for all 6
  months
• Only 1 large object lived in the top 10 for 6
  months
• 44 large files remained in the top 100 for 6
  months
• The most popular small objects are new objects
• Most requests are for old objects

10
Measurement, Modeling, and Analysis
of a Peer-2-Peer File-Sharing Workload

• Fetch-at-most-once
• Once a KaZaA user obtains an object, they will
  not need to retrieve another copy
• 94 of Objects are fetched once per user
• 99 are fetched less than twice per user
• Stems from the fact that media files are
  immutable and never stale
• You may refresh slashdot.org three times a day,
  but there is no point download thriller.mpeg
  seventeen times.
• This keeps KaZaA workload from following a Zipf
  curve even though object popularity does.

11
Measurement, Modeling, and Analysis
of a Peer-2-Peer File-Sharing Workload

• Workload Modeling
• Create a set of objects and give them popularity
  based on a zipf distribution
• Create a set of clients that requests objects in
  proportion to there popularity
• Have each client fetch-at-most-once
• Measure the distribution of transfers
• Does it follow a zipf curve
• How many big-object requests can a population of
  size N satisfy

12
Measurement, Modeling, and Analysis
of a Peer-2-Peer File-Sharing Workload
Popular objects are not requested as curve would
predict
13
Measurement, Modeling, and Analysis
of a Peer-2-Peer File-Sharing Workload

• Would a proxy cache help?
• At first the proxy will cache the popular objects
  and succeed.

• But as fetch-at-most-once draws clients away
  from the Zipf curve and the proxy begins to fail.
• What happens if we increase density of
  popularity?
• Curve starts higher and falls faster

14
Measurement, Modeling, and Analysis
of a Peer-2-Peer File-Sharing Workload

• Previous model did not insert new objects.
• New popular objects tend to correct the work
  load.

• Through providing locality
• New clients however do not help, they contribute
  to keeping old objects popular and destroy
  locality

15
Measurement, Modeling, and Analysis
of a Peer-2-Peer File-Sharing Workload

• Validating The Model
• Capture parameters that are inputs to the model
  from the trace
• Number of clients
• Number of objects
• User request rate
• Probability user requests given file - Guess
• Probability of popularity of new objects - Guess
• Object arrival rates Guess
• Run simulation with harvested parameters
• See if simulation predicts what actually happened

16
Measurement, Modeling, and Analysis
of a Peer-2-Peer File-Sharing Workload
Simulation seems to successfully predict reality.
But with three free variables used to tune
results, is this fair?
17
Measurement, Modeling, and Analysis
of a Peer-2-Peer File-Sharing Workload

• What inefficiencies can we eliminate?
• Analysis against the trace shows
• 86 of object transfers were from external
  sources when an internal source possessed the
  object.
• A traditional proxy, given the resources, could
  cut bandwidth utilization by 86
• Would have to host pirated data
• Could use a proxy redirector instead. Must know
  the availability of the objects
• Control traffic is obfuscated
• Build locality into the protocol
• Does this sacrifice anonymity?

18
Measurement, Modeling, and Analysis
of a Peer-2-Peer File-Sharing Workload

• How successful would a locality aware protocol
  be?
• Assume that a client is available for periods the
  trace shows it as active
• During a file transfer - extremely conservative

19
Measurement, Modeling, and Analysis
of a Peer-2-Peer File-Sharing Workload

• Questions?
• Will increasing efficiency decrease load as the
  authors would like? Or simply increase work
  achieved per dollar? Do clients have insatiable
  appetites?
• Are you worried that a large number of queries
  might have already been locally satisfied?