Introduction

Introduction

• acknowledge dangers, pitfalls, limitations of internet simulations
• measurement of internet required as a “reality check”
• experiments vital for dealing with implementation issues
• simulations are good for preventing a “success disaster”
  • the propagation of software across the internet which is poorly (or could be better-designed) which results in suboptimal performance
• simulations are good for examining particular aspects of proposed changes
  • no single suite that encompasses everything
• simulations not good for face values numbers (e.g. speed of protocol A vs B)

An Immense Moving Target

• internet is growing exponentially YoY
• number of bytes being moved is also changing
• exercise caution in assumptions about observations made at a particular point drawing meaning to another point in time

Heterogeneity

• topology and link properties
  • in simulation: which topology to use? how are links connected? properties of links?
  • topology is constantly changing
  • engineered by competing entities
  • properties of links are known, but the parameters span a large range
    • modems w/ hundreds of bytes/s vs fiber links
  • routing changes regularly on the internet – simulation probably needs to include “multi-pathing”
• protocol differences
  • different implementations of the same protocol (TCP) results in different connection performance
• traffic generation
  • need many, many sources to mimic internet traffic
  • could use internet traces to get good representation – but in the real world with congestion, clients cut send rates to reduce bandwidth during congestion
  • often application traces are more useful in creating a realistic scenario than a packet-driven framework
  • not all sources can be characterized by traffic traces
    • user behavior changes depending on traffic conditions. Need to have a deeper understanding of traffic to model this
  • atypical drop rates is somewhat common. Modeling this behavior is not understood.

Today’s Network is not Tomorrow’s

• areas of change
  • pricing structure
  • scheduling
  • wireless
  • impoverished devices (handhelds)
  • web caching
  • multicast
  • “killer apps”

Coping Strategies

• invariants: some property that holds empirically for a wide range of scenarios
  • e.g.:
    • diurnal activity patterns
    • self-similarity: long term correlations in packet arrivals are described as “fractal” or “self-similar”
    • poisson session arrivals
    • log-normal connection sizes: logarithm of sizes is approximated well with gaussian distribution
    • heavy-tailed distributions
      • characterizing network activity, expect to find heavy tails (high variance)
    • invariant distribution for telnet packet generation
      • distribution of network packets generated by telnet session is invariant –> keystrokes can probably be modeled as an exponential distribution
    • invariant characteristics of global topology:
      • there are 7 continents, speed of light, distance between NY/Paris, etc