\documentclass{article}
\usepackage{twocol}
\usepackage{amssymb}
\usepackage{epsfig,psfig}

\setlength{\columnsep}{8mm}
\setlength{\oddsidemargin}{-0.25in}
\setlength{\evensidemargin}{-0.25in}
\setlength{\topmargin}{-12mm}
\setlength{\textheight}{222mm}
\setlength{\textwidth}{177mm}


%\newcommand{\doublespace}{\renewcommand{\baselinestretch}{2}\large\normalsize}
%\newcommand{\singlespace}{\renewcommand{\baselinestretch}{1}\large\normalsize}
%\newcommand{\halfspace}{\renewcommand{\baselinestretch}{1.5}\large\normalsize}

\pagestyle{empty}
\makeatletter
\def\section{\@startsection {section}{1}{\z@}{-3.5ex plus -1ex minus
 -.2ex}{2.3ex plus .2ex}{\normalsize\bf}}
\makeatother
\makeatletter
\def\subsection{\@startsection {subsection}{1}{\z@}{-3.5ex plus -1ex minus
 -.2ex}{2.3ex plus .2ex}{\normalsize\bf}}
% put a period after section or subsection number in header
\def\@sect#1#2#3#4#5#6[#7]#8{\ifnum #2>\c@secnumdepth
     \def\@svsec{}\else
     \refstepcounter{#1}\edef\@svsec{\csname the#1\endcsname.\hskip .75em }\fi
     \@tempskipa #5\relax
      \ifdim \@tempskipa>\z@
        \begingroup #6\relax
          \@hangfrom{\hskip #3\relax\@svsec}{\interlinepenalty \@M #8\par}%
        \endgroup
       \csname #1mark\endcsname{#7}\addcontentsline
         {toc}{#1}{\ifnum #2>\c@secnumdepth \else
                      \protect\numberline{\csname the#1\endcsname}\fi
                    #7}\else
        \def\@svsechd{#6\hskip #3\@svsec #8\csname #1mark\endcsname
                      {#7}\addcontentsline
                           {toc}{#1}{\ifnum #2>\c@secnumdepth \else
                             \protect\numberline{\csname the#1\endcsname}\fi
                       #7}}\fi
     \@xsect{#5}}
% put a period after theorem and theorem-like numbers
\def\@begintheorem#1#2{\it \trivlist \item[\hskip \labelsep{\bf #1\ #2.}]}
\makeatother

\title{THE CURRENT STATE AND LIKELY EVOLUTION OF THE INTERNET}
\author{Andrew Odlyzko \\
AT\&T Labs \\
Florham Park, NJ 07932, USA \\
amo@research.att.com \\
http://www.research.att.com/$\sim$amo
}
\begin{document}
\maketitle

\section*{Abstract}
Surprisingly little is known about the Internet.  Even such
basic facts as the size of the networks that make up the
Internet or the amount of traffic they carry are not available.

This paper presents estimates of the main
statistics about the size and growth of the Internet, as well
as about utilization patterns.  This data is then used to justify
some speculative predictions
about the likely evolution of data networks.

\section*{1. Introduction}
This paper presents some of the highlights of the studies
of data networks that are documented in
\cite{CoffmanO,FishburnO,Odlyzko1,Odlyzko2,Odlyzko3}
and in a few cases updates them.  Much more detail about
methodologies and results is available in those papers.
This paper and those studies consider only high-level aggregate measurements
of the Internet, and do not look at details of protocols,
say.  

There are many studies of the
economics of the Internet.  Most of them are listed
in \cite{MacKieM,McKnightB,Varian}.  However, they are
old (by Internet standards) and
none of them answer such basic questions
as how large the various parts of the Internet are,
and how much they cost.

A key point of the investigation of
\cite{CoffmanO,FishburnO,Odlyzko1,Odlyzko2,Odlyzko3}
was the need to
consider not just the public Internet, but the full universe
of data networks and their role in the economy.  For simplicity,
only networks inside the U.S. were considered.  Since costs
of transmission are much lower in the U.S. than in most other
countries, these networks are likely to reflect the behavior
of the Internet in other parts of the world as costs come down.

Even in the restricted realm of data networks, the public
Internet (those parts of the Internet accessible to general
users) is only a fraction, although a noticeable and
rapidly growing fraction, of the total system.  Measuring
networks by their maximal transmission capacity, it was
estimated in \cite{CoffmanO} that at the end of 1997 in the U.S., the switched
voice network was probably still the largest, but the private line
networks were about as large, and the public Internet was
considerably smaller.  More recent updates of the
estimates of \cite{CoffmanO}, using the same methodology,
show the following estimates for the end of 1998.
(The bandwidth of data networks in the table is
the effective bandwidth, as defined in
\cite{CoffmanO}, which is about half of the sum of bandwidths
of all links.  This measure was introduced to compensate for
most packets traveling over about two links, as well as for
data links being shorter than voice links.  See \cite{CoffmanO}
for the detailed justification.)

\begin{center}
\begin{tabular}{lr}
network & bandwidth (Gbps) \\ \hline
US voice & 375 \\
public Internet & 150  \\
other public data networks & 80  \\
private line & 400 \\
\end{tabular}
\end{center}
Thus looking just at the public Internet does not give a proper
perspective on data networks, especially since utilization patterns
of private networks are considerably different, as will be explained
below.

Although data networks are about as large as the voice
network in bandwidth, the voice network still dominates
in carried load, and is likely to do so for a few more years.
The traffic, measured in TB/month (terabytes per month),
through various networks at the end of 1998 is estimated
to have been (in another update of \cite{CoffmanO}):

\begin{center}
\begin{tabular}{ll}
network & traffic (TB/month) \\ \hline
US voice  & 43,000 \\
public Internet & 5,000 - 8,000 \\
other public data networks & 1000  \\
private line & 4,000 - 7,000  \\
\end{tabular}
\end{center}
A comparison of the two tables above shows that there
are substantial differences in utilization rates between
the voice network and data networks.  These differences
can be used to infer what user preferences in data
services are, and how much they are willing to pay.
The basic argument (others are discussed later and
in the papers mentioned before) is that low utilization
rates show that what matters to users is the peak bandwidth,
the ability to carry out transactions quickly, and
not the ability to send many bits.  That is clearly
what is driving the development of local area networks,
and the evidence cited in this paper shows that most
long distance data networks behave that way.

Section 2 is devoted to disproving a variety of common
myths about the Internet.
Section 3 presents some speculations about the
evolution of the Internet.

\section*{2. Common wisdom or common misconceptions?}
Much of the ``folk knowledge'' about the Internet
is simply false.  This section discusses the most important
examples.

{\em Traffic on the Internet is ``only'' doubling every year.}
Many press accounts, even in the professional data networking world,
continue to claim that traffic on the Internet doubles every
three to four months, corresponding to annual growth rates
of 700\% to 1,500\%.  The paper \cite{CoffmanO} showed that
traffic on Internet backbones did grow about 1,000\%
in each of 1995 and 1996.  We do have fairly reliable statistics
for traffic at the end of 1994, when most of it was on the
NSF backbone, and also for the end of 1996, when most is thought
to have passed through the public peering points, for which
data is available.  Thus the high growth rates for 1995 and 1996
appear to be trustworthy.  Less complete data appeared to
show that by 1997 growth had slowed down
to about 100\% a year \cite{CoffmanO}.  That is a remarkably high growth
rate, but nowhere near as high as claimed in the popular
press accounts.  An update of \cite{CoffmanO} showed that
this same growth rate of about 100\% seemed to hold also in 1998.
These estimates are not precise, and the true growth rate
could be 80\% or 120\%, but it is almost certainly well below
200\%, much less the ``doubling every three months'' that
is sometimes cited.

{\em Packet networks are not necessarily more efficient than the switched
voice network.}
In general publications it is often asserted without
qualification that packet networks are less expensive than
the switched voice network.
Some of the new packet-only carriers have
been claiming that IP transport saves more than 90\%
over the cost of traditional switched networks.
In particular, savings on transport costs are widely perceived
as the main advantages of carrying voice over packet networks.
On the other hand, when one considers existing corporate networks,
and compares total costs and the volume of traffic, then 
it appears \cite{Odlyzko2}
that most corporations spend more on transferring large
files over their internal IP networks than they would if they used
modems over the public switched voice network.  This is an astounding
result, since modems use only a small fraction of the
bandwidth of the digital channel
that is provided for voice calls, and network costs
of voice calls are small compared to the prices charged.
The estimates for the cost (to corporations, which is not the
same as the cost to the carriers) of transmitting a megabyte of data
over various networks are estimated in \cite{Odlyzko2} as follows:

\begin{center}
\begin{tabular}{ll}
network & dollars/MB \\ \hline
modem & 0.25 - 0.50 \\
private line & 0.50 - 1.00 \\
Frame Relay & 0.30 \\
Internet & 0.04 - 0.15 \\
\end{tabular}
\end{center}
This table suggests an obvious question:
Why don't corporations junk their private networks and send
data via modems over the public switched voice network?
The answer is that the cost estimates of the table apply
only to large file transfers, and do not take into account
other factors, such as latency.  As an example, a credit card authorization
involves transfer of only a few hundred bytes, and so would
cost far more over a modem than the table might suggest.
It would also take far longer, tens of seconds instead of seconds,
and thus lead to lower productivity of the sales force and
customer dissatisfaction.
There are thus unbeatable advantages to packet networks, but they are not
in network costs, but in flexibility.

The above table raises another question, namely why don't
corporations junk their private networks and send
data via the Internet?  This time the primary reason is
the lack of security and high transmission quality on the Internet.
Another reason is inertia, which is shown by the slow transition
from private line networks to Frame Relay, which does provide
the security and high transmission quality that the Internet
lacks.  Frame Relay is growing rapidly, at almost the rate of the Internet,
but is not stopping private lines from continuing to grow.

{\em The public Internet is still small relative to other data networks.}
Although it is the public Internet that has caught all the attention,
it is still dwarfed in transmission capacity and especially in
costs by the private
line networks, as was shown in the tables in the Introduction.  It is also far
smaller than the switched voice network.  However, it is growing
much faster, about 100\% a year, than either the voice network,
which is growing at around 10\% per year, or the private line
networks, which are growing at around 20-30\% per year.  (See \cite{CoffmanO}
for details.  The information about the sizes of the private line
networks is derived from published accounts by consulting companies
that specialize in collecting such data, such as Vertical Systems.  
The estimates about the
size of Internet backbones were assembled from a variety of sources,
primarily network maps published by ISPs, usually accessible
through the online Boardwatch directory at 
$\langle$http://www.boardwatch.com$\rangle$.)
Therefore if current growth rates continue, then
in a few years the public Internet will
be the dominant communication network, but it is not that yet.

{\em Few data networks are congested.}
A surprising fact is that even though it provides high quality
service, the switched voice network has considerably higher average utilization
than any large collection of data networks.
There is a general perception that the public Internet is hopelessly
crowded, and even most network experts believe
that private line networks are congested as well.  Reality is
different, as is shown in \cite{Odlyzko1} and summarized in the table
below.
(The utilization rates in the table above, and elsewhere
in this paper, refer to averages over a full week.  Busy hour
averages are higher, of course.  For example, for private line
networks, the busiest hour of a business day typically sees
utilization of 15-25\% of capacity, whereas for the voice
network the corresponding figure is around 70\%.  However, it is long term
averages that point out most drastically the different
utilization patterns of the various networks and suggest
ways to improve economics of data transport.)

\begin{center}
\begin{tabular}{lc}
network & utilization \\ \hline
local phone line & 4\% \\
U.S. long distance switched voice & 33\% \\
Internet backbones & 10-15\% \\
private line networks & 3-5\% \\
LANs & 1\%
\end{tabular}
\end{center}
The low utilization of data networks is a key finding
that underlies most of the interpretations and speculations
that follow later in this paper.
This finding was extremely controversial when it
was first publicized during the summer of 1998
in the preprint \cite{Odlyzko1}
(the first work to study this question systematically),
and it is still not
universally accepted.  Particularly suspect was the
claim that the Internet backbones were utilized at
half or even one third the rate of the voice network.
However, there is now more data available supporting
these estimates.  For example, AboveNet, a substantial
ISP with a national backbone and even a trans-Atlantic
link, makes detailed statistics for its network
publicly available (at $\langle$http://www.above.net/traffic$\rangle$).
In the first half of 1999 these statistics showed that the
utilization of AboveNet's long-haul backbone was running
around 16\%.

The estimates of network utilization in \cite{Odlyzko1} were
based on extensive data for a variety of networks.  Still,
even those studies leave much to be desired.  In particular,
data about utilization of private line networks is scarce,
although they form the bulk of all data networks.  That
is why the private line entry in the table spans nearly
a factor of two.
For details of the data, see \cite{Odlyzko1}.

Some parts of the Internet are highly congested,
especially the public peering points, the NAPs and MAEs.
Many university links to the public Internet are also
heavily loaded, which may have persuaded generations of
students that all networks are heavily utilized.
However, the
backbones of the Internet are relatively lightly loaded.
The estimates of their utilization rates in \cite{Odlyzko1}
(based partially on estimates of sizes of various networks
and the traffic they carry
in \cite{CoffmanO}) are consistent
with recent measurements which show
that as long as transmission stays on a single backbone,
latency and jitter are not a problem.  What is congested
are many of the feeder links to the backbones from smaller
ISPs, especially those that aggregate modem traffic.
Fig. 1 shows the traffic pattern on a T1 line (1.5 Mbps)
belonging to an ISP.
\begin{figure}
\centerline{\psfig{file=Pzoc5,width=3.26in,height=2in}}
\caption{Traffic on an ISP's T1 line on Tuesdays of April 14 and 21, 1998.
5-minute averages.}
\end{figure}
It runs at a high fraction of its
capacity for large parts of the day, but still manages
to provide relatively high quality service, with minor
delays and packet losses, according to the network
manager in charge of that link.  (Average utilization is in the
40-45\% range.)  Other ISP links show even higher utilization,
frequent saturation, and high packet loss rates.
(Other examples of traffic patterns of ISPs,
as well as those of other users, are in \cite{Odlyzko1,Odlyzko2,Odlyzko3}.)

As the tables in the Introduction show, most of the data
transmission capacity is in private corporate networks.
Their traffic patterns tend to be far different from those
of the ISP line profiled in Fig. 1.
Fig. 2 shows
utilization of a corporate T1 line.
\begin{figure}
\centerline{\psfig{file=Pzoc4,width=3.26in,height=3.5in,height=2in}}
\caption{Traffic on a corporate T1 line in the continental U.S. during
Thursday, May 28, 1998.  5-minute averages.}
\end{figure}
The average utilization
of this line is slightly under 1\%.  Comparing the graphs of Figures 1
and 2, it is easy to grasp that the performance of those
two lines will be different, and that traffic control algorithms
suitable for one might not fit the other one.

While most corporate networks are run at low average
utilizations, there are many exceptions.  The most prominent
are international lines, such as the one profiled in Fig. 3.
\begin{figure}
\centerline{\psfig{file=Pnortel1,width=3.26in,height=3.5in,height=2in}}
\caption{Traffic from the U.S. to the Far East on a corporate
128 Kbps line during a weekday.  The peak traffic hours between
1800 and 2400 coincide with the busy hours in the Far East location.
Hourly averages.}
\end{figure}
The average traffic shown there is 59 Kbps, or 46\% of capacity
during the day that is profiled.
On this particular link there is little traffic in the
reverse direction, so average utilization of the entire line
(which, as is always the case in current data network as a legacy
of the switched voice network, consists of two one-directional links),
during a business
day is around 25\%.  Over a full week, average utilization
might therefore be expected to be around 20\%.
However, according to the network managers in charge of the line,
it does experience high packet loss rate (in excess of 25\%)
during peak traffic
periods, and provides low quality transmission.

{\em Congestion is not necessarily the biggest problem on the Internet.}
The ``World Wide Wait'' is often caused by problems other than
lack of bandwidth.  A study by 
Huitema \cite{Huitema} about accessing some popular servers showed
that 20\% were not reachable.  Among the 80\% that could be
reached, 42\% of the delays were caused by network transmission, with DNS
accounting for 13\% and servers for the remaining 45\%.
Further,
there are some indications that in the last year, the performance
of the backbones has improved, while servers are falling behind.

{\em ``The tragedy of the commons'' may not be an insurmountable threat for
the Internet.}
It is widely believed that queueing by congestion is how the
Internet is run right now, and that this will not change until
usage sensitive pricing is introduced, 
since demand, driven by flat rate
pricing, is insatiable \cite{GuptaSW}.
However, in a dynamic environment with
growing bandwidth, this argument is questionable.
In many cases, growth has been orderly.  Fig. 4
shows the average traffic from the public Internet to the University
of Waterloo.  (See \cite{CoffmanO,Odlyzko2} for more details.)
\begin{figure}
\centerline{\psfig{file=Pwat2,width=3.26in,height=3.5in,height=2in}}
\caption{Traffic on the link from the public Internet to the University
of Waterloo.  The line with circles shows average traffic during
the month of heaviest traffic in each school term.  The step
function is the full capacity of the link.
By permission of University of Waterloo.}
\end{figure}
Although the capacity
of the link has had several sudden jumps, usage has grown at a
pretty steady 100\% a year.  Similar steady growth rates have been seen in other
networks, see \cite{CoffmanO}.
Thus these networks have not in general had to cope
with sudden surges in demand that saturated new capacity as soon
as it became available.  Even when such surges materialized
(as they did at the University of Waterloo when student dorms
were hooked up to the campus Ethernet), they were contained by
simple local measures, primarily quotas on traffic to individual
PCs.

Not only is growth of data traffic steady, actual traffic is generally
predictable once it is sufficiently aggregated.  Several of the graphs
in this paper, as well as many of those in \cite{Odlyzko1,Odlyzko2}
combine displays of traffic for
several days.  It is noteworthy that the traffic patterns are
generally consistent from week to week, with Monday through Thursday
usually behaving the same, and Friday, Saturday, and Sunday each having
its own particular load graph.  This is the same behavior that has
been observed on the switched voice network.

{\em There are many inefficiencies in data networks that are not
being exploited.}  Most attention is currently devoted to
Quality of Service (QoS) measures \cite{FergusonH}, but without
providing quantitative estimates of how much such measures will
save, or will improve the quality of transmission.  However, there are many
other steps that can be taken to provide a better Internet, steps whose
benefits can often be quantified much more easily and reliably.
For example, costs could be lowered if utilization were
increased by combining corporate private line traffic on
public networks, using Virtual Private Networks (VPNs).
A major reason for the high utilization rate of the switched
voice network is that it carries traffic from both business
and residential customers, and those two classes of users
have complementary traffic patterns, as is shown in Fig. 5.
\begin{figure}
\centerline{\psfig{file=Patt5,width=3.26in,height=3.5in,height=2in}}
\caption{Residential (thin line) and business (line with circles) voice traffic
on U.S. long distance switched voice networks, as percentage of peak traffic
on those networks.}
\end{figure}

Historically there has been considerable asymmetry in public
Internet traffic between the U.S.
and Europe and Asia, with the U.S. sending more bytes than
it receives.  This asymmetry has been increasing in the last
couple of years, so that on many links the ratio is 2:1 or
even 3:1.  (For example, on the British JANET network, in March
1997, 3.73 TB were received from the U.S, and 2.95 were sent
there, for a ratio of of 1.26.  In March 1999, the corresponding
figures were 19.52 and 9.51, for a ratio of 2.05.  On the Swiss
SWITCH network, during the month ending on Feb. 4, 1999, the
corresponding traffic figures were 3.34 and 1.29, for a ratio
of 2.59.)
Hence one could obtain much better capacity
utilization by building transmission systems that do not have
the symmetric links that are a heritage of the switched
voice network.  The gain from doing this is easy to quantify,
unlike the potential gains from implementing many of the
current QoS measures, for which there seem to be no
hard numerical figures.

{\em The ``bursty nature of data traffic'' is not the culprit behind
low utilization rates of data networks.}
Data traffic does not smooth out as well as switched voice traffic,
and it shows long range dependence \cite{FeldmannGWK,LelandTWW}.
However, that does not
mean that high utilization cannot be achieved.  Figures 3 and 6
show that it can.  In Fig. 6, we see essentially full
utilization over 9 hours during the business day, and in
Fig. 3, for a much smaller link, more than 80\% utilization over
a comparable period.  (The ``goodput,'' or measure of 
traffic that end users care about, ignoring the retransmissions,
is presumably much lower, but no estimates for it are available.)
Most of data transport uses TCP, which fills available bandwidth
and can produce high load factors.
Thus low utilization has to come from a different source.
\begin{figure}
\centerline{\psfig{file=Pswitch2,width=3.26in,height=3.5in,height=2in}}
\caption{Traffic on the 8 Mbps link from the U.S. to SWITCH, the Swiss
academic and research network, during Tuesdays of February 3, 10, and 17,
1998.  Hourly averages, Swiss time.  By permission of SWITCH.}
\end{figure}


The paper \cite{Odlyzko1} identified a variety of reasons
why utilization rates of data networks are likely to lag
behind those of the voice network.  These reasons include
rapid growth rate, asymmetry in data traffic, and the much
lower prices per unit of bandwidth of higher capacity links.
These reasons by themselves explain most of the difference
in utilization patterns of ISP networks and the voice network.
Private line networks have the additional disadvantage that
they carry traffic primarily during the business day, and
thus lose the advantage of having complementary traffic
patterns that help keep ISP and voice system pipes full.
Still, that does not fully explain the low utilization
rates of private line networks.

In some cases there is a clear rationale for the design of
data networks.  When large files are to be backed up to
an off-site facility, the networks are sized appropriately
to that task, with some margin of safety, and show moderate
utilization rates.  In other cases, say in online transaction
processing, there are stringent requirements for how long
a transaction can take, and networks are designed accordingly,
usually resulting in low utilization.  In most cases, though,
there is no clear rationale, and designers use a variety of
``rules-of-thumb,'' such as not allowing the utilization
of a T1 to exceed 50\% over more than a certain fraction
of 5-minute intervals during a business day.  Ultimately
such rules appear to come from subjective judgements of
the end-users.  Looking at utilization rates and utilization
patterns, it appears that the main driving force in the
development of data networks is
the desire for low transaction latency.  (This is my
interpretation of the data, and it has to be said
that it is not universally accepted.)
Customers do not
care about networks as such, only about applications.
Only a few people are consciously aware of what they
are doing with data networks.  As an example of such a person,
the manager of a branch lab of a major software producer,
who has a private
line from that lab to company headquarters, said that

\begin{quote}
  I see peak bandwidth as the basic commodity I buy. ...  When we
  had a 256Kb data line it was too slow (it interfered with productivity).
  With a T1 line, no one has complained.  I guess our T1 line is less
  than 1\% utilized. ...  I would not go for a T3 line (it would not
  improve our productivity) but I would not cut back on the T1 line.
\end{quote}

High bandwidth can to some extent compensate for high packet
latency (cf. \cite{Cavanagh}), since it is the time for the
total transaction (such as a Web page download) that matters
to the user, not the time that the first packet makes it
through.
(In cases of extreme latency,
such as satellite channels, protocols that spoof TCP by
sending false acknowledgements to the server from a gateway
are often employed to allow the full bandwidth to be utilized.)
The main point, though, is
that high bandwidth is absolutely essential for low transaction
latency.
If a 5 MB PowerPoint presentation has to be transmitted
from a telecommuter's home to her office, it will take
over 20 minutes over a 56 Kbps modem (which can transmit
upstream at only around 30 Kbps), but in favorable conditions
under a minute over a good ADSL connection.
That private line networks in the continental U.S. have
low utilization rates shows that the desire for low transaction
latency is the driver when costs are not too high.  That similar
private lines across the oceans are heavily utilized, with high
packet loss rates and similar impairments, shows that when
costs are very high, the end-user desire for low transaction
latency is subordinated to the need to lower costs through
high utilization.

\section*{3. The future of the Internet}
In this section I speculate about the future of the Internet.
These speculations are
based on the facts presented in previous sections and conclusions
that I drew from those facts.

In the debates about the future of the Internet, there are arguments
for preserving a single best-effort service
class without state inside the network,
and with low utilization providing high quality of service
for all traffic.  Such arguments have often been supported
by citations of progress in optical transmission, which
promises much lower costs for data links.  The counterargument
has typically been that no matter how low the cost, there
would always be some cost, and so the service providers
would have an incentive to maximize utilization and therefore
would have congested links.  Another, related counterargument
has been that data networks suffer from ``the tragedy of
the commons,'' and no matter how much transmission capacity
is built, it will quickly fill up (as happens almost universally
with roads) \cite{GuptaSW}.

The observations of the preceding section provide strong
evidence in favor of the hypothesis that one can build a
best-effort stateless backbone network that will offer high quality
transport to all traffic primarily through low utilization.
The argument is more subtle, but also more convincing, than
simply saying that prices of transmission will come down,
and therefore we will be able to afford larger pipes.
The point is that, as was discussed at the end of the
preceding section, what people care about is not transmitting
bits, but transmitting them quickly, to achieve low transaction
latency.  Therefore, if prices of data networks decrease
as fast as technological progress suggests they should, data networks
will likely evolve in the same direction that LANs and
computers have, namely towards low utilization.
Note that the counterarguments cited at the beginning
of this section (about data pipes filling up) ought to apply to LANs and
PCs, yet both of them are very lightly utilized.
Although PCs are not free, new 400 MHz Pentium II machines
are being bought even for secretaries.  Yet old 486
machines could in principle do the job.  Those fast new PCs
are purchased for their peak performance, to load word
processors or to recalculate spreadsheets rapidly.
Their average utilization is immaterial.  Similarly,
10 Mbps LANs are being replaced by 100 Mbps ones, and
100 Mbps LANs are beginning to be replaced with gigabit
LANs not because the older networks could not carry
the traffic that is offered to them, but because these older slower
networks
have high transaction latency.  The evidence about
low utilization of private line networks in the continental
U.S. shows that people want the same properties of their
long distance data links that they demand from their LANs and
their PCs.  Further, in the majority of cases they are already
able to buy this performance by obtaining high bandwidth
links that they use at low rates.  Therefore it seems
likely that as prices of data links decrease, the Internet
will evolve towards lightly utilized links.  (For more
details on this argument, and further ones, see
\cite{Odlyzko3}.)

A caveat that has to be offered is that the conclusion
about the feasibility and desirability of a single best-effort
service class is based on two key assumption.  One is
that prices of data transmission will decrease in line
with progress in photonics.  (As is documented in
\cite{CoffmanO}, but is not widely known, prices in
the U.S. that are paid by corporate network managers and
ISPs that do not own their own physical network
did decrease rapidly during the 1980s, but
have been climbing since 1992.  Recently there have
been some signs of a change, and definite declines
have been reported in other countries that had not
experienced the North American declines of the 1980s,
but there is no general trend of decreasing prices yet.)
The other assumption
is that traffic on the Internet will continue to be
dominated by transactions such as Web-surfing and file
transfers, and
not by real-time video and audio.  The argument for
the second assumption is that while there is likely
to be extensive video and audio traffic, it will be
in the form of file transfers (such as MP3 ones) for
later playback on a variety of information appliances,
and not in streaming form.  However, this is definitely
a hypothesis.  For arguments supporting this hypothesis,
see \cite{Odlyzko3}.  One of those arguments comes from
observing the development of LANs.
Those already are moving towards speeds of 100 Mbps
and above, which are far more than enough to accommodate
streaming media.  This shows that when prices are sufficiently
low, the desire for low transaction latency does produce
high bandwidths all by itself.

\begin{thebibliography}{99}

\bibitem{Cavanagh}
J. P. Cavanagh, {\em Frame Relay Applications: Business and
  Technical Case Studies,} Morgan Kaufman, 1998.

\bibitem{CoffmanO}
K. G. Coffman and A. M. Odlyzko, The size and growth rate of the Internet.
  {\em First Monday,} Oct. 1998, $\langle$http://firstmonday.org/$\rangle$.
  Also available at $\langle$http://www.research.att.com/$\sim$amo$\rangle$.

\bibitem{FeldmannGWK}
A. Feldmann, A. C. Gilbert, W. Willinger, and T. G. Kurtz,
  The changing nature of network traffic: Scaling phenomena,
  {\em Computer Communication Review,} 28, no. 2 (April 1998), pp. 5--29.
  Available at $\langle$http://www.research.att.com/$\sim$agilbert$\rangle$.

\bibitem{FergusonH}
P. Ferguson and G. Huston, {\em Quality of Service: Delivering
  QoS on the Internet and in Corporate Networks,} Wiley, 1998.

\bibitem{FishburnO}
P. C. Fishburn and A. M. Odlyzko,
  Dynamic behavior of differential pricing and Quality of Service options
  for the Internet, pp. 128--139
  in {\em Proc. First Intern. Conf. on Information and
  Computation Economies (ICE-98)}, ACM Press, 1998.   Available
  at $\langle$http://www.research.att.com/$\sim$amo$\rangle$.

\bibitem{GuptaSW}
A. Gupta, D. O. Stahl, and A. B. Whinston, The Internet:  A
future tragedy of the commons?, available at
$\langle$http://cism.bus.utexas.edu/res/wp.html$\rangle$.

\bibitem{Huitema}
C. Huitema, The required steps towards high quality Internet services,
  unpublished Bellcore report, 1997.

\bibitem{LelandTWW}
W. E. Leland, M. S. Taqqu, W. Willinger, and D. V. Wilson,
  On the self-similar nature of Ethernet traffic (extended version),
  {\em IEEE/ACM Trans. Networking} 2 (1994), 1-15.

\bibitem{MacKieM}
J. MacKie-Mason, Telecom Information Resources on the Internet,
  Web site with links to online sources,
  $\langle$http://china.si.umich.edu/telecom/telecom-info.html$\rangle$.

\bibitem{McKnightB}
L. W. McKnight and J. P. Bailey, eds., {\em Internet Economics,}
  MIT Press, 1997.  Preliminary version of many papers available in
  {\em J. Electronic Publishing,} special issue on Internet economics,
  $\langle$http://www.press.umich.edu/jep/$\rangle$.

\bibitem{Odlyzko1}
A. M. Odlyzko, Data networks are lightly utilized, and will stay
  that way.  Available at
  $\langle$http://www.research.att.com/$\sim$amo$\rangle$.

\bibitem{Odlyzko2}
A. M. Odlyzko, The economics of the Internet: Utility, utilization,
  pricing, and Quality of Service.  Available at
  $\langle$http://www.research.att.com/$\sim$amo$\rangle$.

\bibitem{Odlyzko3}
A. M. Odlyzko, The Internet and other networks: Utilization rates
  and their implications.  Available at
  $\langle$http://www.research.att.com/$\sim$amo$\rangle$.

\bibitem{Varian}
H. R. Varian, The economics of the Internet, information
  goods, intellectual property and related issues, reference Web
  pages with links,
  $\langle$http://www.sims.berkeley.edu/resources/infoecon/$\rangle$.

\end{thebibliography}

\end{document}