Internet research Needs Better Models Sally Floyd, Eddie Kohler ISCI Center for Internet Research, Berkeley, California

1
Internet research Needs Better ModelsSally
Floyd, Eddie KohlerISCI Center for Internet
Research, Berkeley, California
Presented by Max Podlesny
2
Outline

• Motivation
• Network Model Principles
• Several problems
• Phase Effects
• Active Queue Management Oscillations
• TCP Variants
• Proposal
• Conclusion

3
Motivation

• Little relationship to Internet reality
• Unknown relationship to Internet reality
• What divergences are acceptable?
• Are used models valid?
• Measurements and methodologies have never been
  synthesized into a convenient , coherent hole

4
Network Model Principles

• The full range of parameters that might affect a
  simulation or experiment, i.e.
• network topology
• traffic generation
• end-node protocol behavior
• queue drop policies
• congestion levels
• etc.

5
Several typical models
6
Requirements to the model

• Model should be specific to the research
  questions being investigated
• Model must go hand-in-hand with measurement
• Model should be applicable both to the Internet
  of the future and to the Internet of present
• How do models parameter settings affect
  experimental results?

7
ExamplePhase Effects

• Sensitive dependence on precise parameter
  settings
• It is not relevant to the modern Internet
• Concrete example S.Floyd, V.Jacobson. On Traffic
  Phase Effects in Packet-Switched Gateways.
  Internetworking Reseacrh and Experience, 3(3),
  Sept.1992
• Two TCP flows sharing a Drop-Tail queue
• Simulation topology is a simple dumbbell
• Long-lived flows
• No reverse-path traffic

8
Results of simulations
9
Real network

• Traffic includes short-lived flows
• Traffic consists of small control packets as well
  as large data packets
• More than two competing flows

10
Example Active Queue ManagementOscillations

• Implicit disagreement about which simulation
  scenarios are the most important to address
• Queue oscillations are considered a serious
  potential problem with RED AQM
• Changes in the traffic mix can affect oscillation
  dynamics

11
Model

• A dumbbell topology with a 15 Mbps
• 10ms congested link with Adaptive RED queue
  management
• Similar, small amounts of reverse-path traffic
• All run for 100 seconds
• Difference
• Traffic mixes
• Flow RTTs

12
Results of simulations
13
Actually used models
14
ExampleTCP Variants

• TCP Reno
• TCP Vegas

15
TCP Reno

• Based on acknowledgements
• Two types of congestion event
• Duplicate acknowledgement
• Timeout
• Works well when only one packet is dropped
• Losses often come in bursts
• The problem is of how to avoid retransmit timeouts

16
TCP Vegas

• Based on packet delays
• Optimized only for environments
• having a few active TCP connections
• Sending rate of a TCP connection affects the
  queue size at the router
• Problems arise with higher level of statistical
  multiplexing

17
Proposal

• Questions around congestion-related mechanisms at
  router queues
• Analysis of the questions is supposed to lead to
  description of experimental parameters relevant
  for constructing models
• Simulations are supposed to show how parameter
  settings affect the observed behavior of existing
  techniques
• For settings affecting behavior, new measurement
  studies and analysis of the measurement
  literature are supposed to describe how the
  settings look on real networks

18
Conclusion

• Network research has a great need for better
  models
• Each specific research problem requires its own
  model
• Base of a model must be network measurement if it
  is necessary
• Model should be applicable to the Internet of
  present, and to the Internet of future
• A better understanding of which aspects of models
  are critical for a particular research issue is
  required

19
Questions?