Difficulties in Simulating the Internet
Sally Floyd, Senior Member, IEEE, and Vern Paxson
IEEE/ACM Transactions on Networking, August 2001
Mingyu Sun
Instructor: Dr. Herman D. Hughes CSE808 Fall 2001
Today’s topics




Nov. 12, 2001
Role of simulations in Internet research
Difficulties in simulating the Internet
Strategies to accommodate
Develop a common network simulator
Mingyu Sun, CSE808 Presentation
2
Role of simulation

Measurement
– Is needed for crucial ‘reality check’
– Challenge implicit assumptions

Experiment
– For dealing with implementation issues
– For understanding behavior of intractable systems

Analysis
– Providing the possibility of exploring a model of the
Internet

Simulation
– Complementary to analysis
Nov. 12, 2001
Mingyu Sun, CSE808 Presentation
3
Measurement and experiment

What can do:
– explore “real world”
– the existing Internet
– new environment

What cannot do:
– explore future Internet
– explore different possible architectures
– “Success Disaster”
Example: HTTP protocol used by World Wide Web
Nov. 12, 2001
Mingyu Sun, CSE808 Presentation
4
Analysis and simulation

Analysis:
– Fundamental role
– Complete control the model of Internet
– Disadvantage: risk of losing Internet key behavior

Simulation:
–
–
–
–
Complementary to Analysis
Check correctness
Overcome the difficult, impossible cases in analysis
Develop research intuition
Nov. 12, 2001
Mingyu Sun, CSE808 Presentation
5
Simulation: dangers and pitfalls






Dynamic aspect of Internet
Picking the underlying models
Models used in Analysis and simulation
Capable, accurate simulators
Community requirement: immediate or long term
Validating simulation: make scripts public
Nov. 12, 2001
Mingyu Sun, CSE808 Presentation
6
Properties of Internet

Unify diverse networking technologies and
administrative domains.
– Uniform connectivity,not uniform behavior

Big Size
– 99.8 million computers at the end of 2000
– Range of heterogeneity is large
– Scaling problem: works fine with small number of
computers, but not on large scale

Changes: drastically
Nov. 12, 2001
Mingyu Sun, CSE808 Presentation
7
USENET Traffic Volume

Nov. 12, 2001
Bytes per day sent through
the USENET bulletin
board system, averaged
over two-week intervals.
The growth rate
corresponds to exponential
increase of 80% per year.
Mingyu Sun, CSE808 Presentation
8
LBNL Traffic Volume

Internet connections per
day at LBNL. The growth
rate corresponds to
exponential increase of
52% per year.
 LBNL- Lawrence
Berkeley National
Laboratory
Nov. 12, 2001
Mingyu Sun, CSE808 Presentation
9
Growth of LBNL’s WWW Traffic

Nov. 12, 2001
World Wide Web (HTTP)
connections per month at
LBNL.The growth rate
corresponds to doubling
every 7- 8 weeks.
Mingyu Sun, CSE808 Presentation
10
Heterogeneity -- Topology and link

Hard to characterize the protocol over a range of
topologies
– Constantly change, secret topological information
– Topology generator based on current Internet
– Large-scale nature of network topology and protocols

Modems Vs. fiber optic links
 Wire (Traverse copper, glass wire) Vs.Wireless (radio,
infrared-based)
 Point-to-point Vs. broadcast links (multi-access)
 Land-based Vs. satellite links( geosynchronous orbit, lowearth orbit)
 Dynamic routing (latency, bandwidth, load of paths)
 Routing asymmetric (large topology, scaling)
Nov. 12, 2001
Mingyu Sun, CSE808 Presentation
11
Heterogeneity -- Protocol


Conceptually, the Internet uses a unified set of protocols
In reality, each protocol has been implemented differently.
– Some studies are sensitive to the details of the protocol
– different user, different features, even bugs
– Example: TCP 400 implementation and versions.TCP Tahoe and
Reno replaced by NewReno and SACK

One must decide which applications to simulate using
those protocols
– Different applications have major different characteristics
– These characteristics can vary considerably from site to site
Nov. 12, 2001
Mingyu Sun, CSE808 Presentation
12
Heterogeneity -- Traffic

Basic question:
– How to introduce different traffic sources into the simulation,
while retaining the end-to-end congestion control

Trace-driven
– Problem: Rate adaptation from end-to-end congestion control
causes shaping
– Example: a connection observed on a high-speed unloaded link
might still send packages at a rate much lower than what the link
could sustain because somewhere else along the path insufficient
resources are available.
– Solution: Trace-driven source-level simulation preferable to tracedriven packet-level because data volume and the application-level
pattern are NOT shaped by the network’s current property
Nov. 12, 2001
Mingyu Sun, CSE808 Presentation
13
Problems with source-level trace-driven

NOT all sources can be reliably characterized by traffic
traces
– Example: remote login users terminate or justify commands
– Applications that are inherently adaptive (Internet Video) will be
shaped at both packet-level and application level.

Complexity of traffic congestion level to the network link
– From all degrees of congestion to none at all
– Predicting future evolution of congestion harder than
characterizing certain level of congestion at a particular time

Scaling issue arises using source models of individual
connections to generate aggregated cross-traffic for
simulation
– Expensive to simulate a traffic aggregate on a per source basis
Nov. 12, 2001
Mingyu Sun, CSE808 Presentation
14
Rapid and unpredictable change

Possibilities for unpredictable areas of change:
–
–
–
–
–
–
–
–

Pricing structures
Scheduling
Wireless
Impoverished devices
Native multicast
Differentiated service
Ubiquitous web-caching
a new “killer app”
Implications:
– Our research should not be heavily biased by network details that
are likely to change.
– Our research should not be invalidated by major architectural
changes that might or might not come to dominate the Internet
architecture several years down the road.
Nov. 12, 2001
Mingyu Sun, CSE808 Presentation
15
Coping strategies (I)
Search for invariants

What is invariant?
– Some facet of behavior that has been empirically shown to hold in a
very wide range of environments (traffic characteristics, call arrival
processes, session durations…)

Promising candidates:
– Diurnal patterns of activity
• Different patterns for different protocols, NNTP
• Different patterns for the same protocols, work/leisure surf
• Geographic effects on across time zone communication
–
–
–
–
–
–
Self- similarity
Poisson session arrivals
Log-normal connection sizes
Heavy-tailed distributions
Invariant distribution for Telnet packet
Invariant characteristics of the global topology
Nov. 12, 2001
Mingyu Sun, CSE808 Presentation
16
Coping strategies (II)
Exploring the parameter space
 Explore network behavior as a function of changing
parameters
– Parameters include protocol specifics, router queue management,
packet scheduling, network topologies and link properties, or
traffic mixes.
– Rule of thumb: Order of magnitude in parameter

Pitfall: Internet includes nonlinear feedback mechanisms
that may cause mistakes made by simulation artifacts not
present in the real world.
– Correct interpretation of simulation result
Nov. 12, 2001
Mingyu Sun, CSE808 Presentation
17
NS simulator

What is NS?
– NS is a multiprotocol simulator that implements unicast and multicast routing
algorithms, transport and session protocols (including both reliable and
unreliable multicast protocols), reservations and integrated services.
– NS incorporates a range of link-layer topologies and scheduling and queue
management algorithms.
– NS also incorporates libraries of network topology generators and traffic
generators, the network animator NAM, an emulation interface to allow the NS
simulator to interact with real-world traffic.

Advantage:
– Ability to build upon the work of others in sharing a common simulator

Disadvantages:
– affected by the same bugs or the same modeling assumptions that subtly skew
the results.
– No simulator eliminates the difficulties inherent in Internet simulation

Trends in network simulation: Parallel and distributed simulators
Nov. 12, 2001
Mingyu Sun, CSE808 Presentation
18
Final notes

Discussed issues & difficulties in modeling
Internet traffic, topology and protocols
 Acknowledged limitations & potential of
simulations and model-based research
 Interpreting simulation results and drawing
conclusions
 Besides intuition and good judgment, rely on other
tools (measurements, experiments and analysis)
Thank you!
Nov. 12, 2001
Mingyu Sun, CSE808 Presentation
19