Internet Growth Trends
Dr. Lawrence G. RobertsJanuary, 2000
The first node of the Internet was installed in Sept 1969. My grand plan for the network at that time was to connect 15 computers across the US as an experiment. However, the experiment kept growing and we had connected 52 computers by the time I left ARPA in four years. For 18 years Internet hosts doubled every 15 months and the network traffic doubled every 12 months. Then, in 1997, after DWDM started cutting communication costs in half every 12 months, the market responded by doubling the traffic every 6 months, reflective of a market elasticity of two. Simultaneously, the advent of e-commerce fueled the growth of the traffic. One result of this extremely high growth rate (4 x per year) is that the maximum speed of core routers/switches must increase at the same rate, the first time in history that improvements have been required faster than the improvement rate for semiconductors, Moore’s Law. This trend is likely to continue until about 2008 when DWDM will have used up all the bandwidth in the fiber. Then, unless another rapid decrease in the cost of bandwidth occurs, traffic increase should slow down.
In 1969, two years before Moore published his Law of semiconductor performance increase doubling every 18 months, I published a study ( http://www.ziplink.net/~lroberts/Forecast69.htm ) of the performance increase of 38 computers released or planned to be released) from 1958 to 1972. This study looked at the bits per second that the CPU could process divided by the price of the computer as the performance and did a mean square fit of an exponential to this data. At the time the result was a trend of throughput per dollar of computers doubling every 18.6 months. However, updating the data with current 1999 computers (PC’s), the trend over 41 years is a doubling of computer performance every 21 months.
A similar trend for packet switches from the first ARPA IMP in 1969 to the most modern routers and ATM switches in the 1990-1999 period, confirms that packet switches have followed the same trend as computers with a trend of performance per dollar doubling every 21 months.
Communication Cost TrendIn the years when packet switching was emerging, the cost performance of the leased lines required to make a cross country packet network were decreasing so slowly that they only halved every 79 months. This was during the telecommunications monopoly period and well into the partially competitive period from 1960 to 1995. As a result of this very slow decrease in communications cost and the rapid decrease in the cost of computing, it became economic in 1961 to add computing in the form of packet switching to a communications network to divide data traffic into packets, switch these packets, and statistically concentrate to fill the trunks at each node rather than waste bandwidth but use far less computing with circuit switching. This was because data has a 15:1 peak to average utilization if circuit switched. For voice, with only a 3:1 peak to average utilization, the crossover was in 1969 although, with only a 3:1 gain possible, packet switched voice has not found an economic market until recently as the Internet became larger than the voice network and compressed voice has become more interesting.
In the mid 1990’s Dense Wave Division Multiplexing(DWDM) made it possible to use different colors in the same fiber to multiply the capacity. This has resulted in a major decrease in the cost of long-haul communications estimated at a factor of two every 12 months. This shift to where communications cost is decreasing faster than computing cost has resulted in the possibility that circuit switching might again become economically attractive if the cost of statistically concentrating at a node were more than the cost of the bandwidth that would be wasted. This concept has led many people to predict that optical circuit switching will replace packet switching at Internet core nodes. This however, is not destined to occur since before this can happen (about 2008), the total capacity of the fiber (about 100 Tbps) will be reached and the gains due to DWDM will end. Computing will still be les expensive than fiber capacity, and statistical concentration at each node will still be economic.
Trunk Speed Trend
In 1969 the maximum speed trunk available for computer communications was 50 Kbps. Trunk line speed slowly increased over the years to 10 Gbps today. The electronics to modulate wires and now fibers have been the limiting factor since the final signal must be serial. The trend of the trunk speed increase is a doubling of the speed every 22 months. This is almost identical to the computer performance increase of 21 months, probably as a result of the same driving force; density changes in semiconductors.
Internet Traffic Trend
In 1969 the Internet traffic was very small based on the average traffic in the peak busy hour. As host computers were added and the host protocol was implemented from 1970 to 1982, the traffic grew at about the same rate as computer performance, doubling every 21 months. In 1983, with the transition to TCP/IP complete, the traffic started doubling every 9 months. Then in 1998 the rate increased to doubling every 6 months. This latest trend of four times increase per year is consistent with the reported market elasticity of 2:1 and the factor of two per year decrease caused by the DWDM impact on communications cost. This would suggest that the current traffic trend will last until the DWDM gains are completed in 2008 and then go back to something more like doubling every 8 months if the elasticity remains 2:1 and the communications cost trend goes to a 20 month doubling. However, many factors are at work causing the current traffic growth besides market elasticity and the future traffic growth is very hard to predict except that it will stay high for many years.
Voice Traffic Trend
Voice has stayed very flat for many decades. Cell phones and rapidly decreasing cost will certainly change the amount of time each person uses voice communication, but the overall impact versus data will most likely be minor.Data traffic and voice traffic are currently about equal with voice revenues estimated as 80% of the total revenue. This means that by 2002 with data doubling every 6 months, data will be 94% of the traffic and 80% of the revenue. Only a small fraction of the voice will be packetized by 2002 so that very little price change will have occurred on average. Thus, by the time voice volume really increases due to elasticity, voice will be a minor factor overall.
Maximum Router/Switch Speed Required in Large Cities
In 1969 the Internet used routers with a capacity of four trunks or 200 Kbps. The traffic load was much less than 200 Kbps but the speed was required in order to achieve the packet delay and peak throughput desired. As trunk lines increased in speed, the throughput of Internet routers was increased commensurately. Generally the routers and later ATM switches were designed to support a small number of these maximum speed trunks and a larger number of smaller speed lines. This pattern continued until 1997 when the traffic caught up with the router/switch design. After 1997, the throughput of routers or switches in the large cities needed to be wire speed and support more and more maximum speed trunks. The throughput requirement now must grow at the same rate as the Internet traffic is growing, four times per year.
The relationship between total Internet traffic and the router/switch speed required in the large cities is quite simple. Internet traffic makes 15 hops today and the number of hops increases slowly as the log of the number of nodes. All traffic entering the net thus must be switched 16 times. The largest cities in the world each have 4.1% of the traffic. Internet trunks generally can only be loaded to 50% due to the fractal nature of the traffic. Thus, if the largest carrier has 33.3% of the traffic, then his largest city will require a router switch with a throughput of (.333*16*.041/.50) or 44% of the total traffic. This factor of about 2 between the total traffic and the router/switch throughput will stay relatively constant over time and trend toward one as the hop count increases. On the above graph with a log scale, the total traffic and the router/switch throughput therefore stay very close together after 1997.
Moore’s Law describes the performance increase for semiconductors and was estimated in 1971 as an 18 month doubling. If the last 30 years experience were used to correct this rate it is possible that the 18 months would be slightly different, like 21 months, but clearly this trend has continued with little change. Computer performance is strongly influenced by the semiconductor trend but need not be identical. The very accurate result I have obtained for computer performance is very similar to Moore’s Law but not identical as might be expected. The trend for trunk speed was far less likely to be identical to semiconductor performance but somehow is very close to the computer performance trend. These trends and the communication cost trends are all important to understanding the trends of Internet cost/bit and then via elasticity, the current Internet traffic trend. The traffic trend then predicts the router/switch speed required.
Trend Doubling Period NameSemiconductor performance18 months Moore’s Law
Computer performance/dollar 21 months Roberts Law
Communications- bits/dollar before 1995 79 months
Communications- bits/dollar with DWDM 12 months
Maximum Internet Trunk Speed in service 22 months
Internet Traffic Growth 1969-1982 21 months
Internet Traffic Growth 1983-1997 9 months
Internet Traffic Growth 1997-2008 6 months
Internet Router/Switch Max Speed until 1997 22 months
Internet Router/Switch Max Speed after 1997 6 months
Copyright © 2001 Dr. Lawrence G. Roberts