`Windy City Innovations, LLC, Patent Owner 1
`
`
`
`1038
`
`Paxson and Floyd
`
`administered by vastly different policies to seamlessly
`interoperate. However, the fact that IP masks these
`differences from a user’s perspective does not make
`them go awayl IP buys uniform connectivity in the
`face of diversity, not uniform behavior.
`A second key property is that the Internet is big.
`The most recent estimate is that it included more
`
`than 16 million computers in Jan. 1997 (Letter,
`1997). Its size brings with it two difficulties. The first
`is that the range of heterogeneity mentioned above is
`very large:
`if only a small fraction of the computers
`behave in an atypical fashion, the Internet still might
`include thousands of such computers, often too many
`to dismiss as negligible.
`the crucial problem of
`Size also brings with it
`scaling: many networking protocols and mechanisms
`work fine for small networks of tens or hundreds
`of computers, or even perhaps “large” networks of
`10’s of 1,000’s of computers, yet become impractical
`when the network is again three orders of magnitude
`larger (and likely to be five orders of magnitude larger
`within a decade). Large scale means that rare events
`will routinely occur in some part of the network, and,
`furthermore, that reliance on human intervention to
`maintain critical network properties such as stability
`becomes a recipe for disaster.
`A third key property is that the Internet changes
`in drastic ways over time.
`For example, we men-
`tioned above that in Jan. 1997, the network included
`16 million computers.
`If we step back a year in
`time to Jan. 1996, however, then we find it included
`“only” 9 million computers. This observation then
`begs the question: how big will it be in another year?
`3 years? 5 years? One might be tempted to dismiss its
`near—doubling during 1996 as surely a one—time phe—
`nomenon. However, this is not the case. For example,
`Paxson (1994a) discusses measurements showing sus—
`tained Internet traffic growth of 80% /year going back
`to 1.984. Accordingly, we cannot assume that the net—
`work’s current, fairly immense size indicates that its
`growth must surely begin to slow.
`Unfortunately, growth over time is not the only way
`in which the Internet is a moving target. Even what
`we would assume must certainly be solid, unchanging
`statistical properties can change in a brief amount of
`time. For example, in Oct. 1992 the median size of
`an Internet FTP (file transfer) connection observed at
`LBNL was 4,500 bytes (Paxson, 1994b). The median
`is considered a highly robust statistic, one immune to
`outliers (unlike the mean, for example), and in this
`case was computed over 60,000 samples. Surely this
`statistic should give us some solid predictive power
`in forecasting future FTPl connection characteristics!
`Yet only five months later, the same statistic com-
`
`less
`puted over 80,000 samples yielded 2,100 bytes,
`than half what was observed before. Thus, we must
`exercise great caution in assuming that observations
`made at a particular point in time tell us much about
`properties at other points in time.
`For Internet engineering, however, the growth in
`size and change in connection characteristics in some
`sense pale when compared to another way in which
`the Internet is a moving target: it is subject to major
`changes in how it is used, with new applications some—
`times virtually exploding on the scene and rapidly
`altering the lay of the land. For example, for a re—
`search site studied by Paxson (1994a), the Web was
`essentially unknown until late 1992 (and other traf—
`fic dominated the site). Then, a stunning pattern of
`growth set in: the site’s Web traffic began to double
`every sir weeks, and continued to do so for two full
`years. Clearly, any predictions of the shape of future
`trafic made before 1993 were hopelessly off the mark
`by 1994, when Web traffic wholly dominated the site’s
`activities.
`Furthermore, such explosive growth was not a one
`time event associated with the paradigm—shift in In—
`ternet use introduced by the Web. For example, in
`Jan. 1992 the MBone 77a “multicast backbone” used
`to transmit audio and video over the Internet )did
`
`not exist. Three years later, it made up 20% of all of
`the Internet data bytes at one research lab; 40% at
`another; and more than 50% at a third. It too, like
`the Web, had exploded. In this case, however, the ex—
`plosion abated, and today MBone traffic is overshad—
`owed by Web traffic. How this will look tomorrow,
`however, is anyone’s guess.
`In summary:
`the Internet’s technical and admin—
`istrative diversity, sustained growth over time, and
`immense variations over time in which applications
`are used and in what fashion, all present immense
`difficulties for attempts to simulate it with a goal of
`obtaining “general” results.
`
`3 HETEROGENEITY ANY WHICH WAY
`YOU LOOK
`
`Even if we fix our interest to a single point Of time, the
`Internet remains immensely heterogeneous.
`In the
`previous section we discussed this problem in high-
`level terms; here, we discuss more specific areas in
`which ignoring heterogeneity can completely under-
`mine the strength of simulation results.
`
`3.1 Topology and Link Properties
`
`A basic question for a network simulation is what
`topology to use for the network being simulated—the
`
`2
`
`
`
`Why We Don’t Know How to Simulate the Internet
`
`1039
`
`specifics of how the computers in the network are con—
`nected (directly or indirectly) with each other, and
`the preperties of the links that foster the intercon—
`nection.
`
`Unfortunately, the topology of the Internet is dif—
`ficult to characterize. First, it is constantly chang—
`ing. Second, the topology is engineered by a number
`of different entities, not all of whom are willing to
`provide topological information. Because there is no
`such thing as a “typical” Internet topology, simula—
`tions exploring protocols that are sensitive to topo—
`logical structure can at best hope to characterize how
`the protocol performs in a range of topologies.
`On the plus side,
`the research community has
`made significant advances in developing topology-
`generators for Internet simulations (Calvert, Doar
`and Zegura, 1997). Several of the topology genera-
`tors can create networks with locality and hierarchy
`loosely based on the structure of the current Internet.
`The next problem is that while the properties of
`the different types of links used in the network are
`generally known, they span a very large range. Some
`are slow modems, capable of moving only hundreds
`of bytes per second, while others are state-of—the—
`art fiber optic links with bandwidths a million times
`faster. Some are “point—to—point” links that directly
`connect two computers (this form of link is widely as—
`sumed in simulation studies); others are “broadcast”
`links that directly connect a large number of comput—
`ers (these are quite common in practice).
`Another type of link is that provided by connec—
`tions to satellites. If a satellite is in geosynchronous
`orbit, then the latency up to and back down from the
`satellite will be on the order of 100’s of milliseconds,
`much higher than for most land—based links. On the
`other hand, if the satellite is in low-earth orbit, the
`latency is quite a bit smaller, but changes with time
`as the satellite crosses the face of the earth.
`Another facet of topology easy to overlook is that
`of dynamic routing.
`In the Internet, routes through
`the network can change on time scales of seconds to
`days (Paxson, 1996), and hence the topology is not
`fixed. If the route changes occur on fine enough time
`scales (per-packet changes are not unknown), then
`one must refine the notion of “topology” to include
`“multi-pathing.” Multi—pathing immediately brings
`other complications: the latency, bandwidth and load
`of the different paths through the network might dif-
`fer considerably.
`Finally, routes are often asymmetric, with the route
`from computer A to computer B through the network
`differing in the hops it visits from the reverse route
`from B to A. Routing asymmetry can lead to asym-
`metry in path properties such as bandwidth (which
`
`can also arise from other mechanisms). An interest-
`ing facet of asymmetry is that it often only arises in
`large topologies:
`it provides a good example of how
`scaling can lead to unanticipated problems.
`
`3.2 Protocol Differences
`
`these topology and link property
`Once all of
`headaches have been sorted out, the researcher con-
`ducting a simulation study must
`then tackle the
`specifics of the protocols used in the study.
`For
`some studies, simplified versions of the relevant In—
`ternet protocols may work fine, But for other stud-
`ies that are sensitive to the details of the protocols
`(it can be hard to tell these from the former!), the
`researcher faces some hard choices. While conceptu-
`ally the Internet uses a unified set of protocols, in
`reality each protocol has been implemented by many
`different communities, often with significantly differ—
`ent features (and of course bugs). For example, the
`widely used Transmission Control Protocol (TCP)
`has undergone major evolutionary changes. A study
`of eleven different TCP implementations found distin~
`guishing differences among nearly all of them (Pax—
`son, 1997), and major problems with several.
`Consequently, researchers must decide which real-
`world features and peculiarities to include in their
`study, and which can be safely ignored. For some sim-
`ulation scenarios, the choice between these is clear;
`for others, determining what can be ignored can
`present considerable difficulties.
`After deciding which specific Internet protocols to
`use, they must then decide which applications to sim-
`ulate using those protocols. Unfortunately, different
`applications have major differences in their character-
`istics; worse, these characteristics vary considerably
`from site to site, as does the “mix” of which appli-
`cations are predominantly used at a site. Again, re-
`searchers are faced with hard decisions about how to
`keep their simulations tractable without oversimpli-
`fying their results to the point of uselessness.
`
`4 TRAFFIC GENERATION
`
`For many Internet simulations, a basic problem is how
`to introduce different traffic sources into the simula-
`tion. The difficulty with synthesizing such traffic is
`that no solid, abstract description of Internet traf-
`fic exists. At best, some (but not all) of the salient
`characteristics of such traffic have been described in
`abstract terms, a point we return to in §6.1.
`Trace-driven simulation might appear at first to
`provide a cure-all for the heterogeneity and “real—
`world warts and all” problems that undermine ab—
`
`3
`
`
`
`1040
`
`Paxson and Floyd
`
`If only one
`stract descriptions of Internet traffic.
`could collect enough diverse traces, one could in prin-
`ciple capture the full diversity. This vision fails for a
`basic, often unappreeiated reason. One crucial prop-
`erty of much of the traffic in the Internet is that
`it uses adaptive congestion control. That is, each
`source transmitting data over the network inspects
`the progress of the data transfer so far, and if it de—
`tects signs that the network is under stress, it cuts
`the rate at which it sends data,
`in order to do its
`part in diminishing the stress (Jacobson, 1988). Con—
`sequently,
`the timing of a connection’s packets as
`recorded in a trace intimately reflects the conditions
`in the network at the time the connection occurred.
`Furthermore, these conditions are not readily deter-
`mined by inspecting the trace. Connections adapt
`to network congestion anywhere along the end—t0-
`end path between the sender and the receiver. So
`a connection observed on a high—speed, unloaded link
`might still send its packets at a rate much lower than
`what the link could sustain, because somewhere else
`along the path insufficient resources are available for
`allowing the connection to proceed faster.
`In this paper we will refer to this phenomenon as
`traffic shaping. Traffic shaping leads to a danger-
`ous pitfall when simulating the Internet, namely the
`temptation to use trace-driven simulation to incorpo—
`rate the diverse real—world effects seen in the network.
`The key point is that, due to rate adaptation, we can—
`not safely reuse a trace of a connection’s packets in
`another context, because the connection would not
`have behaved the same way in the new context!
`Traffic shaping does not mean that, from a simu»
`lation perspective, measuring traffic is fruitless.
`In~
`stead, we must shift our thinking away from trace-
`driven packet-level simulation and instead to trace-
`driven source—level simulation. That is, for most ap-
`plications, the volumes of data sent by the endpoints,
`and often the application-level pattern in which data
`is sent (request/reply patterns, for example), is not
`shaped by the network’s current properties; only the
`lower—level specifics of exactly which packets are sent
`when are shaped. Thus, if we take care to use traf—
`fic traces to characterize source behavior, rather than
`packctelevel behavior, we can then use the source
`level descriptions in simulations to synthesize plausi—
`ble traffic. See Danzig et a1. (1992), Paxson (1994b),
`and Cunha, Bestavros and Crovella (1995).
`An alternative approach to deriving source models
`from traffic traces is to characterize traffic sources
`in more abstract terms, such as using many data
`transfers of a fixed size or type. The Internet’s per-
`vasive heterogeneity raises the question: which set
`of abstractions should be used?
`Is the traffic of
`
`the aggregate
`for example,
`interest dominated by,
`of thousands of small connections (Web “mice”), or
`a few extremely large, one-way, rate-adapting bulk
`transfers (“elephants”), or long-running, high-volume
`video streams “multicasted’7 from one sender to mul—
`tiple destinations, or bidirectional multimedia traffic
`generated by interactive gaming?
`A final dimension that must be explored is: to what
`level should the traffic congest the network links?
`Virtually all degrees of congestion, including none at
`all, are observed with non—negligible probability in
`the Internet. More generally, variants on the above
`scenarios all occur in different situations frequently
`enough that they cannot be dismissed out of hand.
`
`5 TODAY’S NETWORK IS NOT TOMOR-
`ROW’S
`
`A harder issue to address in a simulation study con-
`cerns how the Internet might evolve in the future.
`For example, all of the following might or might not
`happen within the next year or two:
`a New pricing structures are set in place, leading
`users to become much more sensitive to the type and
`quantity of traffic they send and receive.
`a The Internet routers switch from their present
`“first come, first serve” scheduling algorithm for ser—
`vicing packets to methods that attempt
`to more
`cquably share resources among difierent connections
`(such as Fair Queueing, discussed by Deniers, Keshav
`and Shenker, 1990).
`0 “Native” multicast becomes widely deployed.
`Presently, Internet multicast is built on top of unicast
`“tunnels,” so the links traversed by multicast traffic
`are considerably different from those that would be
`taken if multicast were directly supported in the heart
`of the network. And/or: the level of multicast audio
`and video traffic explodes, as it appeared poised to
`do a few years ago.
`0 The Internet deploys mechanisms for supporting
`different classes and qualities of service (Zhang et al.,
`1993). These mechanisms would then lead to dilfer~
`ent connections attaining potentially much different
`performance than they presently do, with little inter—
`action between traflic from different classes.
`- Web—caching becomes ubiquitous. For many pur-
`poses, Internet traffic today is dominated by World
`Wide Web connections. Presently, most Web con-
`tent is available from only a single place (server) in
`the network, or at most a few such places, which
`means that most Internet Web connections are “wide-
`area,” traversing geographically and topologically
`large paths through the network. There is great inter—
`est in reducing this traffic by widespread deployment
`
`4
`
`
`
`Why We Don’t Know How to SimuIate the Internet
`
`1041
`
`of mechanisms to support caching copies of Web con-
`tent at numerous locations throughout the network.
`As these efforts progress, we could find a large shift
`in the Internet’s dominant traffic patterns towards
`higher locality and less stress of the wide-area infras-
`tructure.
`
`- A new “killer application” comes along. While
`Web traffic dominates today, it is vital not to then
`make the easy assumption that it will continue to
`do so tomorrow. There are many possible new ap—
`plications that could take its place (and surely some
`unforeseen ones, as was the Web only a few years
`ago), and these could greatly alter how the network
`tends to be used. One example sometimes overlooked
`by serious—minded researchers is that of multi—player
`gaming: applications in which perhaps thousands or
`millions of people use the network to jointly entertain
`themselves by entering into intricate (and bandwidth—
`hungry) “virtual realities.”
`Obviously, some of these changes will have no effect
`on some simulation scenarios. But one often does not
`know a prion." which can be ignored, so careful re-
`searchers must conduct a preliminary analysis of how
`these and other possible changes might undermine the
`relevance of their simulation results.
`
`6 COPING STRATEGIES
`
`So far we have focused our attention on the various
`factors that make Internet simulation a demanding
`and difficult endeavor. In this section we discuss some
`strategies for coping with these difficulties.
`
`6.1 The Search for Invariants
`
`The first observation we make is that, when faced
`with a world in which seemingly everything changes
`beneath us, any “invariant” we can discover then be-
`comes a rare piece of bedrock on which we can then
`attempt to build. By the term’invariant we mean
`some facet of Internet behavior which has been em-
`pirically shown to hold in a very wide range of envi—
`ronments.
`
`Thinking about Internet properties in terms of in—
`variants has received considerable informal attention,
`but to our knowledge has not been addressed system-
`atically. We therefore undertake here to catalog what
`we believe are promising candidates:
`0 Longer—term correlations in the packet arrivals
`seen in aggregated Internet traflic are well described
`in terms of “self-sirnilar” (fractal) processes. To those
`versed in traditional network theory,
`this invariant
`might appear highly counter—intuitive. The standard
`modeling framework, often termed Poisson or Marko—
`
`vian modeling, predicts that longer-term correlations
`should rapidly die out, and consequently that traf—
`fic observed on large time scales should appear quite
`smooth. Nevertheless, a wide body of empirical data
`argues strongly that these correlations remain non-
`negligible over a large range of time scales (Leland
`et al., 1994; Paxson and Floyd, 1995; Crovella and
`Bestavros, 1996).
`time scales
`roughly,
`“Longer-term” here means,
`from hundreds of milliseconds to tens of minutes. On
`shorter time scales, effects due to the network trans-
`port protocols—Which impart a great deal of struc—
`ture on the timing of consecutive packets—are be—
`lieved to dominate traffic correlations, although this
`property has not been definitively established. On
`longer time scales, non—stationary effects such as di-
`urnal traffic load patterns become significant.
`In principle, self—similar traffic correlations can lead
`to drastic reductions in the effectiveness of deploy—
`ing “buffers” in Internet routers in order to absorb
`transient increases in traffic load (Erramilli, Narayan
`and Willinger, 1996). However, we must note that
`the network research community remains divided on
`the practical impact of self-similarity (Grossglauser
`and Bolot, 1996). That self-similarity is still find-
`ing its final place in network modeling means that
`a diligent researcher conducting Internet simulations
`must not a priori assume that its effects can be ig—
`nored, but must instead consider how to incorporate
`self-similarity into any traffic models used in a simu—
`lation.
`
`Unfortunately, accurate synthesis of self—similar
`traffic remains an open problem. A number of al—
`gorithms exist for synthesizing exact or approximate
`sample paths for different forms of self—similar pro—
`cesses. These, however, solve only one part of the
`problem, namely how to generate a specific instance
`of a set of longer-term traffic correlations. The next
`step—how to go from the pure correlational structure,
`expressed in terms of a time series of packet arrivals
`per unit time, to the details of exactly when within
`each unit of time each individual packet arriveswhas
`not been solved. Even once addressed, we still face
`the difficulties of packet—level simulation vs. source-
`level simulation discussed in §4.
`In this regard, we
`note that Willinger et a1. (1995) discuss one promising
`approach for unifying link-level self-similarity with
`specific source behavior, based on sources that ex-
`hibit ON/ OFF patterns with durations drawn from
`distributions with heavy tails (see below).
`0 Network user “session” arrivals are well-described
`using Poisson processes. A user session arrival corre-
`sponds to the time when a human decides to use the
`network for a specific task. Examples are remote lo—
`
`5
`
`
`
`1042
`
`Paxson and Floyd
`
`gins and the initiation of a file transfer (FTP) dialog.
`Unlike the packet arrivals discussed above, which con-
`cern when individual packets appear, session arrivals
`are at a much higher level; each session will typically
`result in the exchange of hundreds of packets. Paxson
`and Floyd (1995) examined different network arrival
`processes and found solid evidence supporting the use
`of Poisson processes for user session arrivals, provid-
`ing that the rate of the Poisson process is allowed to
`vary on an hourly basis.
`(The hourly rate adjust—
`ment relates to another invariant, namely the ubiqui—
`tous presence of daily and weekly patterns in network
`traffic.) They also found that slightly finer-scale ar—
`rivals, namely the individual network “connections”
`that comprise each session, are not well described as
`Poisson, so for these we still lack a good invariant on
`which to build.
`
`0 A good rule of thumb for a distributional fam-
`ily for describing connection sizes or durations is log-
`normal. Paxson (1994b) examined random variables
`associated with measured connection sizes and dura—
`tions and found that, for a number of different ap-
`plications, using a log—normal with mean and vari-
`ance fitted to the measurements generally describes
`the distribution as well as previously recorded empir—
`ical distributions. This finding is beneficial because
`it means that by using an analytic description, we do
`not sacrifice significant accuracy over using an empir-
`ical description; but, on the other hand, the finding is
`less than satisfying because Paxson also found that in
`a number of cases, the fit of neither model (analytic
`or empirical) was particularly good, due to the large
`variations in connection characteristics from site-to-
`site and over time.
`
`0 When characterizing distributions associated
`with network activity, expect to find heavy tails. By
`a heavy tail, we mean a Pareto distribution with
`shape parameter a < 2. These tails are surprising
`because the corresponding Pareto distribution has in—
`finite variance. (Some statisticians argue that infinite
`variance is an inherently slippery propertyihow can
`it ever be verified? But then, independence can never
`be proven in the physical world, either, and few have
`difficulty accepting its use in modeling.) However, the
`evidence for heavy tails is very widespread, including
`CPU time consumed by Unix processes; sizes of Unix
`files, compressed video frames, and World Wide Web
`items; and bursts of Ethernet and FTP activity.
`0 Finally, Danzig and colleagues (1992) found that
`the pattern of network packets generated by a user
`typing at a keyboard has an invariant distribution.
`Subsequently, Paxson and Floyd (1995) confirmed
`this finding, and identified the distribution as hav-
`ing both a Pareto upper tail and a Pareto body.
`
`6.2 Carefully Exploring the Parameter Space
`
`Another fundamental coping strategy is to judiciously
`explore the parameter space relevant to the simula-
`tion. Because the Internet is such a heterogeneous
`world, the results of a single simulation based on a
`single set of parameters are useful for only one thing,
`namely determining whether the simulated scenario
`exhibits a show—stopping problem. As one Internet
`researcher has put it, “If you run a single simulation,
`and produce a single set of numbers (e.g., throughput,
`delay, loss), and think that that single set of numbers
`shows that your algorithm is a good one, then you
`haven’t a clue."
`Instead, one must analyze the re-
`sults of simulations for a wide range of parameters.
`Selecting the parameters and determining the
`range of values through which to step them form an—
`other challenging problem. The basic approach is to
`hold all parameters (protocol specifics, how routers
`manage their queues and schedule packets for for—
`warding, network topologies and link properties, traf—
`fic mixes, congestion levels) fixed except for one ele-
`ment, to gauge the sensitivity of the simulation sce—
`nario to the single changed variable. One rule of
`thumb is to consider orders of magnitude in param—
`eter ranges (since many Internet properties are ob—
`served to span several orders of magnitude). In ad—
`dition, because the Internet includes non-linear feed-
`back mechanisms, and often subtle coupling between
`its different elements, sometimes even a slight change
`in a parameter can completely change numerical re-
`sults (see Floyd and Jacobson, 1992, for example).
`In its simplest form, this approach serves only to
`identify elements to which a simulation scenario is
`sensitive. Finding that the simulation results do not
`change as the parameter is varied does not provide a
`definitive result, since it could be that with other val—
`ues for the fixed parameters, the results would indeed
`change. However, careful examination of why we ob—
`serve the changes we do may in turn lead to insights
`into fundamental couplings between different param-
`eters and the network’s behavior. These insights in
`turn can give rise to new invariants, or perhaps “sim-
`ulation scenario invariants,” namely properties that,
`while not invariant over Internet traffic in general, are
`invariant over an interesting subset of Internet traffic.
`
`7 THE VINT PROJECT
`
`The difficulties with Internet simulation discussed
`above are certainly daunting. In this section we dis—
`cuss a collaborative effort now underway in which a
`number of Internet researchers are attempting to sig-
`nificantly elevate the state-of—the-art in Internet sim—
`
`6
`
`
`
`Why We Don’t Know How to Simulate the Internet
`
`1043
`
`ulationi
`
`The VINT project (http://netweb.usc.edu/vint/)
`is a joint project between USC/ISI, Xerox PARC,
`LBNL, and UCB to build a multi—protocol
`sim—
`ulator, using the UCB/LBNL network simulator
`“ns” (http://www—mash.cs.berkeley.edu/ns/). The
`key goal
`is to facilitate studies of scale and pro-
`tocol interaction. The project will incorporate li-
`braries of network topology generators and traf—
`fic generators, as well as visualization tools such
`as
`the network animator
`“11am”
`(http://www-
`mash .cs.berkeley.edu/ ns/nan1.html) .
`Most network researchers use simulations to ex—
`plore a single protocol. However, protocols at dif-
`ferent layers of the protocol stack can have unan-
`ticipated interactions that are important to discover
`before deployment in the Internet itself. One of the
`goals of the VINT project is building a multi—protocol
`simulator
`that
`implements unicast and multicast
`routing algorithms, transport and session protocols,
`reservations and integrated services, and application—
`level protocols such as HTTP. In addition, the simula—
`tor already incorporates a range of link-layer topolo—
`gies and scheduling and queue management algo-
`rithms. Taken together, these enable research on the -
`interactions between the protocols and the mecha-
`nisms at the various network layers.
`A central goal of the project is to study protocol
`interaction and behavior at significantly larger scales
`that those common in network research simulations
`today. The emphasis on how to deal with scaling
`is not on parallel simulation techniques to speed up
`the simulations (though these would be helpful, too),
`but to allow the researcher to use difiercnt levels of
`abstraction for different elements in the simulation.
`
`The hope is that VINT can be used by a wide range
`of researchers to share research and simulation re—
`sults and to build on each other’s work. This has al-
`ready begun for some areas of Internet research, with
`researchers posting ns simulation scripts to mailing
`lists to illustrate their points during discussions.
`If
`this lingua franca extends to other areas, then VINT
`might play an important role in abetting and system—
`atizing Internet research.
`
`8 THE ROLE OF SIMULATION
`
`We finish with some brief comments on the role of
`simulation as a way to explore how the Internet
`behaves, versus experimentation, measurement, and
`analysis. While in some fields the interplay between
`these may be obvious, Internet research introduces
`some unusual additions to these roles.
`
`Clearly, all four of these approaches are needed,
`each playing a key role. All four also have weaknesses.
`Measurement
`is needed for a crucial
`“reality
`check.” It often serves to challenge our implicit as-
`sumptions.
`Indeed, while we personally have con-
`ducted a number of measurement studies, we have
`never conducted one that failed to surprise us in some
`fundamental fashion.
`
`Experiments are crucial for dealing with implemen—
`tation issues, which can at first sound almost trivial,
`but often wind up introducing unforeseen complex—
`ities. Experimentation also plays a key role in ex-
`ploring new environments before finalizing how the
`Internet protocols should operate in those environ—
`ments.
`
`One problem Internet research suffers from that
`most other fields do not is the possibility of a “suc—
`cess disaster” —designing some new Internet function—
`ality 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 software
`can propagate itself to endpoints over the network,
`it is not at all implausible that the new function-
`ality might spread to a million computers within a
`few months. Indeed, the HTTP protocol used by the
`VVOrld Wide Web is a perfect example of a success
`disaster. Had its designers envisioned it in use by vir-
`tually the entire Internetiand had they explored the
`corresponding consequences with experiments, anal-
`ysis or simulation—they would have significantly al-
`tered its design, which in turn would 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. Analysis plays a fundamental role, because
`it brings with it greater understanding of the forces
`at play. It carries with it, however, the serious risk
`of using a model simplified to the point where key
`facets of Internet traffic have been lost, in which case
`any ensuing results are useless (though they may not
`appear to be sol).
`Simulations are complementary to analysis, not
`only by providing a check on the assumptions of the
`model and on the correctness of the analysis, but
`by allowing exploration of complicated scenarios that
`would be either difficult or impossible to analyze.
`(Simulations can also play a vital role in helping re
`searchers to develop intuition.) The complexities of
`Internet t0pologies and traffic, and the central role
`of adaptive congestion control, make simulation the
`most promising tool for addressing many of the ques-
`tions of Internet traffic dynamics. As we have illus—
`trated, there does not exist a single suite of simula-
`
`7
`
`
`
`104/1
`
`Paxson and Floyd
`
`tions suficient to prove that a proposed protocol per—
`forms well in the Internet environment. Instead, sim—
`ulations play the role of examining particular aspects
`of proposed protocols, and, more generally, particular
`facets of Internet behavior.
`
`For some topics, the simplest scenario that illus-
`trates the underlying principles is often the best. In
`this case the researcher can make a conscious decision
`to abstract away all but what are judged to be the
`essential components of the scenario under study. As
`the research community begins to address questions
`of scale, however7 the utility of small, simple simula-
`tion scenarios is reduced, and it becomes more critical
`for researchers to address questions of topolo