`Service
`
`Andrew Odlyzko
`
`AT&T Labs - Research
`amo@research.att.com
`
`July 7, 1998.
`
`Abstract. Can high quality be provided economically for all transmissions on the Internet? Current
`
`work assumes that it cannot, and concentrates on providing differentiated service levels. However,
`
`an examination of patterns of use and economics of data networks suggests that providing enough
`
`bandwidth for uniformly high quality transmission may be practical. If this turns out not to be possible,
`
`only the simplest schemes that require minimal involvement by end users and network administrators
`
`are likely to be accepted. On the other hand, there are substantial inefficiencies in the current data
`
`networks, inefficiencies that can be alleviated even without complicated pricing or network engineering
`
`systems.
`
`1. Introduction
`
`The Internet has traditionally treated all packets equally, and charging has involved only a fixed
`
`monthly fee for the access link to the network. However, there are signs of an imminent change.
`
`There is extensive work on provision of Quality of Service (QoS), with some transmissions getting
`
`preferential treatment. (For a survey of this area and references, see the recent book [FergusonH].)
`
`Differential service will likely require more complicated pricing schemes, which will introduce yet
`
`more complexity.
`
`The motivation behind the work on QoS is the expectation of continued or worsening congestion.
`
`As Ferguson and Huston say (p. 9 of [FergusonH])
`
`...
`
`it sometimes is preferable to simply throw bandwidth at congestion problems. On a
`
`global scale, however, overengineering is considered an economically prohibitive luxury.
`
`Within a well-defined scope of deployment, overengineering can be a cost-effective alter-
`
`native to QoS structures.
`
`The argument of this paper is that overengineering (providing enough capacity to meet peak de-
`
`mands) on a global scale may turn out not to be prohibitively expensive. It may even turn out to be the
`
`GUEST TEK EXHIBIT 1019
`Guest Tek v. Nomadix, IPR2019-00253
`
`
`
`cheapest approach when one considers the costs of QoS solutions for the entire information technolo-
`
`gies (IT) industry.
`
`Overengineering has been traditional in corporate networks. Yet much of the demand for QoS is
`
`coming from corporations. It appears to be based on the expectation that overengineering will not be
`
`feasible in the future. “There's going to come a time when more bandwidth is just not going to be
`
`available... and you'd better be able to manage the bandwidth you have,” according to one network
`
`services manager [JanahTD].
`
`The abandonment of the simple traditional model of the Internet would be a vindication for many
`
`serious scholars who have long argued that usage-sensitive pricing schemes and differential service
`
`would provide for more efficient allocation of resources. (See [McKnightB] for references and surveys
`
`of this work.) The need for usage-sensitive pricing has seemed obvious to many on the general grounds
`
`of “tragedy of the commons”. As Gary Becker, a prominent economist, said recently (in advocating
`
`car tolls to alleviate traffic jams and the costs they impose on the economy [Becker]):
`
`An iron law of economics states that demand always expands beyond the supply of free
`
`goods to cause congestion and queues.
`
`It may indeed be an iron law of economics that demand for free goods will always expand to
`
`exceed supply. The question is, will it do so anytime soon? An iron law of astrophysics states that the
`
`Sun will become a red giant and expand to incinerate the Earth, but we do not worry much about that
`
`event. Furthermore, the law of astrophysics is much better grounded in both observation and theoretical
`
`modeling than the law of economics. For example, consider Table 1 (based on data from tables 12.2
`
`and 18.1 of [FCC]). It shows a dramatic increase in total length of toll calls per line. Such calls are paid
`
`by the minute of use, and their growth was presumably driven largely by decreasing prices, as standard
`
`economic theory predicts. On the other hand, local calls in the U.S. (which are almost universally
`
`not metered, but paid for by a fixed monthly charge, in contrast to many other countries) have stayed
`
`at about 40 minutes per day per line in the last two decades. The increase of over 62% in the total
`
`volume of local calls was accompanied by a corresponding increase in the number of lines. There is
`
`little evidence in this table of that “iron law of economics” that causes demand to exceed supply, and
`
`which, had it applied, surely should have led to continued growth in local calling per line. (There is also
`
`little evidence of the harm that Internet access calls are supposed to be causing to the local telephone
`
`companies. This is not to say there may not have been problems in some localities in California, for
`
`example, or that there won' t be any in the future. However, at least through 1996 the increasing use of
`
`2
`
`
`
`networked computers has not been a problem in aggregate.)
`
`An obvious guess as to why we have stable patterns of voice calls is that people have limited time,
`
`and so, with flat-rate pricing, their demand for local calls had already been satisfied by 1980. However,
`
`that is not what the data in Table 1 shows. While the total volume of local calls went up almost
`
`63% between 1980 and 1996, population increased only 16.5%, so minutes of local calls per person
`
`(including modem and fax calls) increased by 40%. Thus demand for local calls has been growing
`
`vigorously, but it was satisfied by a comparable increase in lines. Families and businesses decided, on
`
`average, to spend more on additional phone lines instead of using more of the “free good” that was
`
`already available. Somewhat analogous phenomena appear to operate in data networking, and may
`
`make it feasible to provide high quality undifferentiated service on the Internet.
`
`In data networks, at first sight there does appear to be extensive evidence for that “iron law of
`
`economics.” Comprehensive statistics are not available, but it appears that Internet traffic has been
`
`doubling each year for at least the last 15 years, with the exception of the two years of 1995 and 1996,
`
`when it appears to have grown by a factor of about 10 each year [CoffmanO]. Almost every data link
`
`that has ever been installed was saturated sooner or later, and usually it was sooner rather than later.
`
`An instructive example is that of the traffic between the University of Waterloo and the Internet, shown
`
`in Fig. 1. The Waterloo connection started out as a 56 Kbps link, was upgraded to 128 Kbps in July
`
`1993, then to 1.5 Mbps in July 1994, and most recently to 5 Mbps in April 1997 [Waterloo]. Based on
`
`current usage trends, this link will be saturated by the end of 1998, and will need to be upgraded, or
`
`else some rationing scheme will have to be imposed. (A partial rationing scheme is already in effect,
`
`since the link is heavily utilized and often saturated during the day.)
`
`The University of Waterloo statistics could be regarded as clinching the case for QoS and usage-
`
`sensitive pricing. They show a consistent pattern of demand expanding to exceed supply. However,
`
`I suggest a different view. The volume of data traffic in Fig. 1 grows at a regular pace, just about
`
`doubling each year. The 12-fold jump in network bandwidth from 128 Kbps to 1.5 Mbps in July 1994
`
`did not cause traffic to jump suddenly by a factor of 12. Instead, it continued to grow at its usual pace.
`
`The students did not go wild, and saturate the link by downloading more pictures. Similarly, statistics
`
`for traffic on the Internet backbones show steady growth, aside from an anomalous period of extremely
`
`rapid increase in 1995 and 1996 [CoffmanO], and the NSFNet backbone in particular had traffic almost
`
`exactly doubling from the beginning to 1991 to the end of 1994. And should an increase in traffic be
`
`wrong? We are on the way to an Information Society, and so in principle we should expect growth in
`
`data traffic.
`
`3
`
`
`
`How much capacity to provide should depend on the value and the price of the service. To decide
`
`what is feasible or desirable, we have to consider the economics of the Internet. Unfortunately, the
`
`available sources (such as those in the book [McKnightB] or those currently available online through
`
`the links at [MacKieM, Varian]) are not adequate. The information they contain is often dated, and it
`
`usually covers only the Internet backbones. However, these backbones are a small part of the entire
`
`data networking universe. Sections 2 to 6 attempt to partially fill the gap in published information about
`
`the economics of the Internet.
`
`Fig. 2 is a sketch of the Internet, with the label “Internet” attached just to the backbones (as the
`
`term is often used). As will be shown in Section 2 (based largely on the companion papers [CoffmanO,
`
`Odlyzko2]), these backbones are far smaller than the aggregate of corporate private line networks,
`
`whether measured in bandwidth or cost (although not necessarily in traffic). (See Table 2 for the sizes
`
`of data networks in the U.S. It is taken from [CoffmanO], and effective bandwidth, explained in that
`
`reference, compensates for most data packets traveling over more than a single link.) The private line
`
`networks, in turn, are dwarfed by the LANs (local area networks) and academic and corporate campus
`
`networks. Most of the pricing and differentiated service schemes that are being considered, though, are
`
`aimed at Internet backbones or private line WAN links. We need to consider how they would interact
`
`with the other data networks and the systems and people those networks serve.
`
`Most of the effort on QoS schemes is based on the assumption of endemic congestion. However,
`
`when we examine the entire Internet, we find that most of it is uncongested. That the LANs are lightly
`
`used has been common knowledge. However, it appears to be widely believed that long distance data
`
`links are heavily utilized. The paper [Odlyzko2] (see Section 3 for a summary) shows that this belief
`
`is incorrect. Even the backbone links are not used all that intensively, and the corporate private line
`
`networks are very lightly utilized. There are some key choke points (primarily the public exchange
`
`points, the NAPs and MAEs, and the international links) that are widely regarded as major contributors
`
`to poor Internet performance, but there is even some dispute about their significance.
`
`(In general,
`
`while there have been numerous studies of the performance of the Internet, some very careful, such as
`
`[Paxson], there is still no consensus as to what causes the poor observed performance.)
`
`What is not in dispute is that a large fraction of the problems that cause complaints from users are
`
`not caused by any deficiencies in transmission. Delays in delivery of email are frequent, but are almost
`
`always caused by mail server problems, as even trans-Atlantic messages do get through expeditiously.
`
`A large fraction of Web-surfing complaints are caused by server overloads or other problems. There
`
`are myriad other problems that arise, such as those concerned with DNS, firewalls, and route flapping.
`
`4
`
`
`
`A key question is whether QoS would help solve those other problems, or would aggravate them, by
`
`making the entire system more complicated, increasing the computational burden on the routers, and
`
`increasing the numbers and lengths of queues.
`
`Many QoS schemes require end-to-end coordination in the network, giving up on the stateless
`
`nature of the Internet, which has been one of its greatest strengths. Essentially all QoS schemes have the
`
`defect that they require extensive involvement by network managers to make them work. However, it is
`
`already a major deficiency of the Internet that, instead of being the dumb network it is often portrayed
`
`as, it requires a huge number of network experts at the edges to make it work [Odlyzko3]. Instead of
`
`throwing hardware and bandwidth at the problem, QoS would require scarce human resources.
`
`The evidence presented in this paper, combined with that of [Odlyzko2], shows that the current
`
`system, irrationally chaotic as it might seem, does work pretty well. There appear to be only a small
`
`number of choke points in the system, which should not be too expensive to eliminate. Further, there
`
`are some obvious inefficiencies in the system that can be exploited. By moving away from private lines
`
`to VPNs (Virtual Private Networks) over the public Internet, one could provide excellent service for
`
`everybody through better use of aggregation of traffic and complementarity of usage patterns. The bulk
`
`of the work on QoS may be unnecessary.
`
`Anania and Solomon wrote a paper in 1988 (which was widely circulated and discussed at that
`
`time, but was only published recently in [AnaniaS]) that took the unorthodox approach of arguing
`
`for a flat-rate approach to broadband pricing. That paper was about pricing of what are now called
`
`ATM services, which have QoS built in, but many of Anania and Solomon's arguments also imply
`
`the desirability of a simple undifferentiated service. My work presents some additional arguments and
`
`extensive evidence of the extent to which the traditional undifferentiated service, flat-price system can
`
`work.
`
`QoS does have a role to play. There will always be local bottlenecks as well as emergency situations
`
`that will require special treatment. Even when local network and server resources are ample, there
`
`will often be need to ration access to scarce human resources, such as technical support personnel.
`
`Even in the network, methods such as Fair Queueing [FergusonH] can be valuable in dealing with
`
`local traffic anomalies, for example. Implementing them would represent a departure from the totally
`
`undifferentiated service model, but a mild one, and one that can be implemented inside the network,
`
`invisible to the users, and without requiring end-to-end coordination in the network. My argument is
`
`that we need to make the network appear as simple as possible to the users, to minimize their costs.
`
`Sections 2 through 12 describe the economics of the Internet. The conclusion is that with some ex-
`
`5
`
`
`
`ceptions, the system does work pretty well as is. There are bottlenecks, but there are also inefficiencies
`
`that can be exploited to eliminate the bottlenecks. Users in general behave sensibly, and although their
`
`demands for bandwidth are growing rapidly, these demands are reasonably regular and predictable. It
`
`appears likely that unit prices for transmission capacity will decline drastically (although total spending
`
`on high bandwidth connections will surely grow), which should make it economically feasible to meet
`
`the growing demand.
`
`It is impossible to predict with any certainty how the Internet will evolve, especially since its evo-
`
`lution depends on many factors, not only basic computing and networking technology and possible
`
`appearance of the proverbial “next killer app,” but also on government regulation and sociology. Still,
`
`some conclusions can be drawn from the study of the current system. The complexity of the entire In-
`
`ternet is already so great, that the greatest imperative should be to keep the system as simple as possible.
`
`The costs of implementing involved QoS or pricing schemes are large and should be avoided. Section
`
`13 outlines three scenarios that appear most likely. One is the continuation of the current flat rate pric-
`
`ing structure with almost uniform best-effort treatment of all packets, and enough bandwidth to provide
`
`high quality transmission. That scenario is likely to materialize if transmission prices decline rapidly
`
`enough. If they don' t, the second scenario might arise, still with flat rate pricing and undifferentiated
`
`service, but with pricing reflecting expected usage of a customer. Finally, if even greater constraints
`
`are needed on traffic, ones that would provide congestion controls, approaches such as the Paris Metro
`
`Pricing (PMP) scheme of [Odlyzko1] might have to be used. PMP is the least intrusive possible usage-
`
`sensitive pricing scheme possible, and my prediction is that if any usage-sensitive pricing is introduced,
`
`it will eventually evolve towards (or degenerate into) PMP. None of these three scenarios would meet
`
`the conventional standards for economic optimality. However, the main conclusion of this paper is that
`
`optimality is unattainable, and we should seek the simplest scheme that works and provides necessary
`
`transmission quality.
`
`2. The Internet and other networks
`
`There are many excellent technical books and journal articles describing the technologies of the
`
`Internet (cf. [Keshav]). There is also a huge literature on how the Internet will change our economy
`
`and society (cf. [Gates]). On the other hand, practically nothing has been published on how the Internet
`
`is used, and how much it costs. It is as if we had shelves full of books telling us how to build internal
`
`combustion engines, and a comparable set of books on the effects of the automobile on suburban sprawl,
`
`income inequality, and other socioeconomic issues, but nothing about how many cars there were, or
`
`6
`
`
`
`how much they cost to operate.
`
`This section attempts to partially fill this gap in the knowledge of economics of data networks, but
`
`the picture it presents can only be a sketchy one. Still, it should help illuminate the major economic
`
`factors that are driving the evolution of the Internet.
`
`The Internet (sometimes called the global Internet) refers to the entire collection of interconnected
`
`networks around the world that share a common addressing scheme. As such, it includes all of the
`
`elements shown in Fig. 2, which is a grossly simplified sketch of the data networking universe. The
`
`element called “Internet” in Fig. 2 is really just the public Internet, the core of the network consisting of
`
`the backbones and associated lines that are accessible to general users. WANs (Wide Area Networks)
`
`consist of some of the clouds in that figure (which are made up of LANs and campus networks) con-
`
`nected via either private line networks, or via public Frame Relay and ATM data networks provided by
`
`telecommunications carriers, or else via the public Internet. Fig. 2 omits many important elements of
`
`the data networking universe, such as regional ISPs.
`
`This paper will concentrate on data networks in North America, primarily in the United States. Just
`
`as in the companion papers [CoffmanO, Odlyzko2], the justification is that most of the spending on
`
`data traffic is in the U.S. [DataComm]. Further, U.S. usage, influenced by lower prices than in most of
`
`the world [ITU], foreshadows what the rest of the world will be doing within a few years, as prices are
`
`reduced.
`
`Data networks do not operate in isolation. To see them in the proper perspective, let us note that total
`
`spending on information technologies (IT) in the U.S. was about $600 billion in 1997, approximately
`
`8% of gross domestic product. The IT sector of the economy is credited with stimulating the high
`
`growth rate of the economy of the last few years, low unemployment, and low inflation [DOC].
`
`Data communications cost about $80 billion in the U.S. in 1997, or 13% of total IT spending,
`
`according to [DataComm]. Table 2 is a brief summary of the statistics on where this spending was
`
`directed, based on the more detailed information in [DataComm] (which also covers the rest of the
`
`world). These statistics show that transmission accounted for only $16 billion, 20% of total for data
`
`communications, and 2.6% of total for all of IT. Thus data lines are a small part of the entire IT picture,
`
`and any scheme that attempts to improve their performance has to be weighed against costs that it might
`
`impose on the rest of the system. It is better to double the spending on transmission than to increase
`
`the average cost of all other IT systems by 3%.
`
`Let us also note that U.S. spending on phone services from telecommunications carriers was around
`
`$200 billion in 1997. Of this total, about $80 billion was for long distance calls, but about $30 billion
`
`7
`
`
`
`was for access charges, paid to the local carriers. Thus it is more appropriate to say that $50 billion was
`
`for long distance services, and $150 billion for local services. In any event, today much more is being
`
`spent on voice phone services than on data. In the future, as broadband services grow, we can expect
`
`the balance to shift towards data. In particular, looking at total communications spending and how it is
`
`still dominated by voice, it is reasonable to expect substantial growth in spending on data transmission.
`
`The core of the Internet, namely the backbones and their access links, is surprisingly inexpensive.
`
`There are many large estimates for total Internet spending, but those are misleading. There were about
`
`20 million residential accounts with online services such as AOL at the end of 1997. At $20/month,
`
`they generated revenues of around $5 billion per year. However, most of that revenue is used to cover
`
`local access costs (the modems, customer service, and marketing expenses of the ISPs). The back-
`
`bones are only a small part of the cost picture for residential customers (cf. [Leida]). In the statistics
`
`of [DataComm] (and of Table 2) they apparently are included in the ”Commercial Internet services”
`
`category, which came to $1.5 billion in 1997. We now derive two other estimates that are both in that
`
`range. According to industry analysts [IDC], MCI's Internet revenues (which include only a small
`
`contribution from residential customers, and are dominated by corporate and regional ISP links to the
`
`MCI network) came to $251 million in 1997 (a 103% increase over 1996), and were running at an
`
`annual rate of $328 million in the last quarter of 1997. Since MCI is estimated to carry between 20
`
`and 30 percent of the backbone traffic, we can estimate total revenues from all backbone operations
`
`between $1.1 and $1.6 billion at an annual rate at the end of 1997. (With revenues doubling each year,
`
`it is not adequate to look at annual statistics.) Yet another, rough way to estimate the costs of Internet
`
`backbones is to take their size, around 2,100 T3 equivalents at the end of 1997 [CoffmanO], and apply
`
`to that the $20,000 per month average cost of a T3 line [VS]. This produces an estimate of about $500
`
`million per year for the main backbone links. When we add some additional costs for the access lines
`
`from carriers' Points of Presence to their backbones (cf. [Leida]), and apply the general estimate that
`
`for large carriers, transmission costs are about half of total costs, we arrive at an estimate of about $1.5
`
`billion for the costs of the core of the Internet. (Costs and revenues are not the same, especially in the
`
`Internet arena, where red ink is plentiful as various players attempt to build market share, but within
`
`the huge uncertainty bounds we are working with, that should not matter much.)
`
`Compared to the Internet backbones, the total cost of private line networks is at least 6 times as
`
`large. Furthermore, the aggregate bandwidth of leased lines is also much greater than of the public
`
`backbones, although the traffic they carry appears to be comparable in volume. (See tables 2 and 5,
`
`taken from [CoffmanO].) This helps explain why the evolution of the Internet is increasingly dominated
`
`8
`
`
`
`by corporate networks.
`
`Just as with switched voice networks, data network costs are dominated by the local part. However,
`
`there is much more heterogeneity in data than in voice.
`
`In the foreseeable future, large academic
`
`and corporate networks are likely to have 2.4 Kbps wireless links along with 14.4 and 28.8 Kbps
`
`modems, megabit xDSL and cable modem links, and gigabit fiber optic cables. In the local campus
`
`wired environment, overengineering with Ethernet, Fast Ethernet, Gigabit Ethernet, and similar tools
`
`appears certain to be the preferred solution. However, there will still be challenges of interconnecting
`
`the other transmission components (whether slower or faster), as well as all the servers and other
`
`equipment that require the bandwidth. Network managers will have a hard time making everything
`
`interoperate satisfactorily even without worrying about QoS.
`
`The Internet backbones are small and inexpensive compared to the rest of the Internet. However,
`
`they are the heart of the Internet, just like a human heart that is small but crucial for the life of the
`
`body. The role of the backbones will likely become even more important in the future as a result of
`
`several related developments. One is that they are being traversed by an increasing fraction of data
`
`traffic. The traditional 80/20 rule, which said that 80% of the traffic stayed inside a local or campus
`
`network is breaking down. Ferguson and Huston [FergusonH] even mention some networks where as
`
`much as 75% of the traffic goes over long distance links. (We do not know how far that traffic goes,
`
`and in particular whether there will continue to be a strong distance dependence in the future. See
`
`[CoffmanO] for a more detailed discussion.) Another reason for the increasingly important role of the
`
`Internet backbones is that they are supplanting private line networks as corporate WAN links, and with
`
`the development of extranets, will be playing a crucial role in the functioning of the whole economy.
`
`Thus there is a reason to worry about the costs of the backbones, as they might become a larger fraction
`
`of the total networking pie. On the other hand, if one can overengineer only one part of the Internet,
`
`then it is best to do it to the core, as without high quality transmission at the core, other parts of the
`
`network will be only be able to offer poor service.
`
`3. Network utilization
`
`Network utilization rates are seldom discussed, yet they are the main factor determining costs of
`
`data services. A line that is used at 5% of capacity costs twice as much per byte of transmitted data as
`
`one whose average utilization rate is 10%.
`
`Although there is no simple relation between the quality perceived by customers and how heavily
`
`their networks are used, the less heavily loaded the network, the better the service. Even the notoriously
`
`9
`
`
`
`congested trans-Atlantic links do appear to provide good performance for applications as demanding
`
`as packet telephony in the early hours of Sunday morning. What this says is that even without any new
`
`QoS technologies, one can provide excellent quality by lowering utilization. Thus the main problem of
`
`the Internet is not a technical one, but an economic one, whether one can afford to have lightly utilized
`
`networks. As an example, the experimental vBNS network, discussed in [Odlyzko2], provides quality
`
`sufficient for even the most demanding applications, but it is expensive (or would be expensive, were it
`
`operated on a commercial basis), running at an average utilization of its links of around 3%.
`
`The main question for the future of the Internet is whether customers are willing to pay for high
`
`quality by having low utilization rates, or whether many links will be congested, with QoS providing
`
`high quality for some select fraction of data transfers. We can find much about the likely evolution
`
`of data networks from observation of usage patterns of existing networks. When we consistently see
`
`lightly utilized links where customers can obtain higher utilization rates and lower costs by switching
`
`to lower capacity lines, we can deduce that they do want high quality data transport, and are willing to
`
`pay for it.
`
`Table 4 shows utilization rates (averaged over a full week) for various networks. It is based on
`
`[Odlyzko2], except for the entry for local phone lines, which is derived from the data in Table 1 (which
`
`is based on [FCC]). A surprising result is that the long distance switched network is by far the most
`
`efficient in terms of utilizing transmission capacity. For most people, an even more surprising feature
`
`of the data is the low utilization rate of private line networks.
`
`The paper [Odlyzko2] discusses the reasons data networks are lightly utilized. Lumpy capacity is a
`
`major one. Rapid and unpredictable growth is another. Small private networks are yet another. Perhaps
`
`the main reason, though, is the bursty nature of data traffic. This traffic is bursty on both the short
`
`and long time scales, and customers do value such bursty transmission. This means that we cannot
`
`reasonably expect data networks to approach the efficiency with which the switched voice network
`
`uses transmission capacity.
`
`Utilization rates can also provide guides to the extent that QoS measures might improve perceived
`
`quality of networking. Practically all Internet users find service much better 5 in the morning than at
`
`noon. However, traffic on the backbones in the early hours of the morning is still about half that during
`
`the peak hours, as is seen in Fig. 10 of this paper and several figures in [Odlyzko2]. Further, traffic
`
`mix does not seem to vary much between trough and peak periods [ThompsonMW]. Therefore we
`
`can conclude that during peak periods, no QoS measure is likely to provide transmissions that have
`
`priorities around the median of all traffic with better service than the current undifferentiated service
`
`10
`
`
`
`provides for all traffic early in the morning. (There could be some improvements in jitter, for example,
`
`but algorithms that do provide such improvements could be used inside the network, invisibly to the
`
`users, to provide similar improvements for undifferentiated service.)
`
`An important point in considering the low utilization rates of data networks is that it is not caused
`
`primarily by inefficiency or incompetence. Customers choose the capacities of their lines, and their
`
`choices tell us what they want and what they are willing to pay for. This point will be treated at greater
`
`length in the following sections.
`
`4. Inefficiency is good (if you can afford it)
`
`Efficiency in utilization of transmission lines or switches should not be the main criterion for
`
`evaluating how good a network is. The crucial question is how well customers' needs are satisfied.
`
`Consider figures 3 and 4, which show utilizations of dial-in modems at Columbia University and
`
`the University of Toronto. (These figures are based on detailed data supplied by those institutions.
`
`Graphs for more recent periods, separated out further by 14.4 and 28.8 Kbps modem pools, can be
`
`found at [Columbia, Toronto].) The average utilizations (over the periods shown in the figures) were
`
`52% at the University of Toronto and 78% at Columbia University. Clearly Columbia was utilizing its
`
`modems more efficiently. Was it providing better service, though? Its modems were completely busy
`
`for more than 12 hours a day. (The slight drops below 100% in the utilization in Fig. 3 are misleading,
`
`since they represent only a little idle time, and are largely the resetting of modems after a session is
`
`terminated.) Clearly much demand is unsatisfied, and there are many frustrated potential users who do
`
`not accomplish their work. Further, the high 78% utilization rate is misleading, since many users are
`
`probably staying online for longer periods than they would if they had assurance they could get a new
`
`connection when they wanted it. Toronto, with a lower utilization rate, managed to accommodate all
`
`demands except for a brief period on Monday night.
`
`The University of Toronto managed to satisfy essentially all demands of its students and faculty
`
`for modem connections and still achieve a 52% utilization rate. This rate is extremely high. There are
`
`many examples of low utilization rates. The family car is typically used around 5% of the time. The
`
`fax in our office or home, the PC on the desktop, and the road we drive on are all designed for peak
`
`usage, and are idle most of the time. We are willing to pay for this inefficiency because the costs are
`
`low compared to the benefits we receive. As costs decrease, we usually accept lower efficiency. For
`
`example, in the early days of computing, programs were written in assembly language. Later, as the
`
`industry advanced, there was a shift towards compiled programs. They typically run at half the speed
`
`11
`
`
`
`of assembly coded versions, but they make software easier to write and more portable. With further
`
`advances in computing power, industry was willing to jump on the Java bandwagon, even though the
`
`early versions of Java typically ran a hundred times slower than compiled programs. With over 99% of
`
`computing cycles devoted to running screen savers, this was a worthwhile tradeoff.
`
`When capital co