.icir.org/floyd/papers/simulate_2001.pdf. It's a snapshot of the page taken as our search engine crawled the Web.
The web site itself may have changed. You can check the current page or check for previous versions at the Internet Archive. Yahoo! is not affiliated with the authors of this page or responsible for its content.
Difculties in Simulating the Internet
Difculties in Simulating the Internet
Sally Floyd and Vern Paxson
�
AT&T Center for Internet Research at ICSI (ACIRI)
Berkeley, CA 94704
USA
http://www.aciri.org/
February, 2001
To appear in IEEE/ACM Transactions on Networking
Abstract
Simulating how the global Internet behaves is an immensely
challenging undertaking because of the networks great het-
erogeneity and rapid change. The heterogeneity ranges from
the individual links that carry the networks trafc, to the
protocols that interoperate over the links, to the mix of dif-
ferent applications used at a site, to the levels of congestion
seen on different links. We discuss two key strategies for
developing meaningful simulations in the face of these dif-
culties: searching for invariants, and judiciously exploring
the simulation parameter space. We nish with a brief look
at a collaborative effort within the research community to
develop a common network simulator.
1
Introduction
Due to the networks complexity, simulation plays a vital
role in attempting to characterize both the behavior of the
current Internet and the possible effects of proposed changes
to its operation. Yet modeling and simulating the Internet
is not an easy task. The goal of this paper is to discuss
some of the issues and difculties in modeling Internet traf-
c, topologies, and protocols. The discussion is not meant as
a call to abandon Internet simulations as an impossible task;
in fact, one of us (Sally) has continued to use simulations as
a key component of her research for many years. Instead,
the purpose is to share insights about some of the dangers
and pitfalls in modeling and simulating the Internet, in or-
der to strengthen the contribution of simulations in network
research. A second purpose is to clearly and explicitly ac-
knowledge the limitations as well as the potential of simula-
tions and of model-based research, so that we do not weaken
our simulations by claiming too much for them.
�
This work was supported by the Director, Ofce of Energy Research,
Ofce of Computational and Technology Research, Mathematical, Informa-
tion, and Computational Sciences Division of the United States Department
of Energy under Contract No. DE-AC03-76SF00098, and by ACIRI. An
earlier version of this paper appeared in the Proceedings of the 1997 Winter
Simulation Conference, Atlanta, GA, 1997.
We begin with the fundamental role of simulation in In-
ternet research (
�
2), and next explore the underlying dif-
culties (
�
3
�
5) rooted in the networks immense heterogene-
ity and the great degree to which it changes over time. We
then discuss some strategies for accommodating these dif-
culties (
�
6). We nish with a brief look at a collaborative
effort within the research community to develop a common
network simulator (
�
7).
2
The Role Of Simulation
While measurement and experimentation provide a means
for exploring the real world, simulation and analysis are
restricted to exploring a constructed, abstracted model of the
world. In some elds the interplay between measurement,
experimentation, simulation, and analysis may be obvious,
but Internet research introduces some unusual additions to
these roles, in part because of the large scale and rapid evo-
lution of the subject area (i.e., the global Internet).
Measurement is needed for a crucial reality check. It
often serves to challenge our implicit assumptions. Indeed,
of the numerous measurement studies we have undertaken,
each has managed to surprise us in some fundamental fash-
ion.
Experiments are frequently vital for dealing with imple-
mentation issueswhich can at rst sound almost trivial, but
often wind up introducing unforeseen complexitiesand for
understanding the behavior of otherwise intractable systems.
Experimentation also plays a key role in exploring new envi-
ronments before nalizing how the Internet protocols should
operate in those environments.
However, measurement and experimentation have limita-
tions in that they can only be used to explore the existing
Internet; while they can be used to explore particular new en-
vironments, they cannot be used to explore different possible
architectures for the future Internet. (There is no instantia-
tion of a future Internet, on the relevant scale and with the
relevant range of future applications, for our measurement
and experimentation.)
1 One problem Internet research suffers, absent from most
other elds, is the possibility of a success disaster
designing some new Internet functionality that, before the
design is fully developed and debugged, escapes into the real
world and multiplies there due to the basic utility of the new
functionality. Because of the extreme speed with which soft-
ware can propagate to endpoints over the network, it is not
at all implausible that the new functionality might spread to
a million computers within a few weeks. Indeed, the HTTP
protocol used by the World Wide Web is a perfect example of
a success disaster. Had its designers envisioned it in use by
the entire Internetand had they explored the corresponding
consequences with analysis or simulationthey might have
signicantly improved its design, which in turn could have
led to a more smoothly operating Internet today.
Analysis provides the possibility of exploring a model of
the Internet over which one has complete control. The role of
analysis is fundamental because it brings with it greater un-
derstanding of the basic forces at play. It carries with it, how-
ever, the risk of using a model simplied to the point where
key facets of Internet behavior have been lost, in which case
any ensuing results could be useless (though they may not
appear to be so!). Even in light of this risk, as scientists we
need to recognize the fundamental role analysis plays in pro-
viding the bedrock on which to build our understanding of
the Internet. Furthermore, while the network is immensely
complex and dauntingly difcult to encompass, we can and
do make progress (often incremental) towards building this
understanding. Finally, we note that much of what we ar-
gue in this paper about difculties with simulation also ap-
ply to difculties with modeling; the core problem of how to
soundly incorporate immense diversity into simulations like-
wise applies to the challenge of trying to devise models with
truly general applicability.
Simulations are complementary to analysis, not only by
providing a check on the correctness of the analysis, but by
allowing exploration of complicated scenarios that would be
either difcult or impossible to analyze. Simulations can also
play a vital role in helping researchers to develop intuition.
In particular, the complexities of Internet topologies and traf-
c, and the central role of adaptive congestion control, make
simulation the most promising tool for addressing many of
the questions about Internet trafc dynamics.
Because simulations often use more complex models than
those that underly analytical results, simulations can be used
to check that simplifying assumptions in the analytical model
have not invalidated the analytical results. However, simula-
tions also generally share some of the same models used in
the analysis, for example, of a simple topology or of a spe-
cic trafc mix. In this case, the agreement between the sim-
ulations and the analysis is not surprising; the agreement be-
tween simulations and analysis does not show that the model
used by the analytical results is in any sense correct.
In this paper we develop the argument that, due to the
heterogeneity and rapid change in the Internet, there does
not exist a single suite of simulation scenarios sufcient to
demonstrate that a proposed protocol or system will perform
well in the future evolving Internet. Instead, simulations play
the more limited role of examining particular aspects of pro-
posed changes or of Internet behavior, and of adding to our
understanding of the underlying dynamics.
For some topics, such as the division of bandwidth among
competing TCP connections with different roundtrip times,
the simplest scenario that illustrates the underlying princi-
ples is often the best. In this case the researcher can make
a conscious decision to abstract away all but the essential
components of the scenario under study. At the same time,
the results illustrated in simple scenarios are stronger if the
researcher shows that the illustrated principle still applies af-
ter adding complexity to the simple scenario by allowing for
various forms of variability known to prevail in real life.
As the research community begins to address questions of
scale, small, simple simulation scenarios become less useful.
It becomes more critical for researchers to address questions
of topology, trafc generation, and multiple layers of proto-
cols, and to p