Vasilios Darlagiannis, EPFL

1
SP1 Measurements and ModelingWP1c Critical
Network Properties for Peer-to-Peer Systems
Performance Evaluation

• Vasilios Darlagiannis, EPFL
• Partners CB, TAU, HUJI, EPFL, KTH

2
Motivation and Objectives

• Hard to evaluate new models for
• P2P systems, CDNs, caching, etc.
• Limits of approaches
• Analysis, simulation, prototypes, emulation
• Objectives
• To summarize the collected Internet topology and
  traffic measurements into a network generator
  model.
• To summarize the key network properties that are
  important from P2P system design and simulation
  point of view.
• To build a network emulation model based on the
  design and simulation which is coherent with the
  observed network characteristics.

3
Big Picture of the Problem
Non-functionalRequirements
Heterogeneous End-devices
How should we deal with this complexity?
4
Case Studies

• Findings on File-sharing applications
• Bandwidth is the most important network property
• Heterogeneous and asymmetrical connections
• Usually unaware of proximity
• Exchanging bigger and bigger files
• Popular files contribute mostly in the traffic
• Findings on MMOGs
• Latency and jitter are the most important network
  properties
• Dropped packets have less important effect than
  expected
• Low bandwidth requirements
• Heterogeneous treatment influences fairness
• Heavy tailed session distributions

5
User Behavior Uptime Distribution

• Uptime and session distribution
• Several measurements performed
• Not identical results but all agree that the
  distribution is heavier than exponential
• Pareto, Lognormal, Weibull, mixtures
• Important factors
• Most of the measurements have been performed on
  File Sharing applications
• Size of exchanged files grows faster than network
  bandwidth
• Relatively inaccurate measurements

6
Summary of the Survey

• Complex problem with several projections to
  encounter
• Application class specific requirements
• Not trivial how to import all the network
  properties in existing network emulators (i.e.
  ModelNet)
• Related areas ALM, CVEs, C/S MMOGs, etc.

7
ModelNet and GODS emulation environment

• ModelNet is a well known emulation environment
• Evaluating real application code
• Reproducible, well-controlled experiments
• Global Observatory for Distributed Systems (GODS)
  is an extension environment allowing ModelNet to
• Dynamically set network parameters
• Observe global knowledge about system state
• Automate several experiments
• Emulate user churn
• Fine tuning, testing, benchmarking

8
Plans for 2007

• Finalize GODS development
• Fine tune the emulation environment to maximize
  its capabilities
• Extract the observed features from the Dimes and
  Etomic nodes infrastructures
• Parameterize a basic emulation model using the
  Evergrow measurements (Dimes, Etomic)
• Evaluate and test the developed P2P algorithms
  (of SP3) using the emulation model