Traffic Patterns: The Internet Is Entirely Unlike The Telephony

The most intriguing feature of Internet traffic is that it seems to grow at an explosive rate (2-5 times a year, by different estimates)... forever. Well, for at least 15 years now. There are no signs that the growth will slow down any time soon.

The exponential Internet growth is fueled by the introduction of newer and more bandwidth-hungry applications, just in time to get the baton from older "flattened out" applications. Actually, that explanation reverses the effect and the cause; there are always applications that demand more bandwidth than is currently available; and the very fact of availability of bandwidth to provide for acceptable performance of an application is exactly what causes its wide acceptance.

There are several "evident" candidates for the next Internet bandwidth hogs, such as Internet telephony, video telephony, video-on-demand, niche broadcasting, data mining agents and global file systems. The very fact of their existence allows us to be confident that the Internet has another 4-5 orders of magnitude of explosive growth left before (and if) the top part of the S-curve is reached.

The exponential growth is, therefore, a strong design constraint. Any design for future Internet backbone technology must be able to sustain the exponential growth. The fact that the rate of Internet traffic growth exceeds the rate of growth of performance of microelectronic devices (Moore's Law, doubling every 2 years) makes it imperative to seek alternative switching technologies that, unlike the present design of IP routers, would not depend on the progress in performance of electronic circuits.

The only existing non-broadcasting electronic communication infrastructure comparable with the Internet in overall size is the telephone network (referred to as "Plain Old Telephone Service", POTS). However, the telephony has essentially only one application, voice transmission. Because it does not have successive waves of new applications, the growth of telephone networks has not been exponential for a long time now.

The basic characteristic of POTS traffic is that it is connection-oriented; i.e. there's always a fixed-bandwidth (duplex 64Kbps) connection for every telephone conversation. The telephone conversations initiated by different people are practically independent; i.e. activity of telephone users can be adequately modeled by a Poisson process. A Poisson process is basically a stream of events ("arrivals") happening at random times independent of each other. A useful property of a Poisson process is that aggregation of several Poisson streams produces a stream with lower variance of inter-arrival times. In other words, the more subscribers a telephone company has, the more the pattern of their calls resembles a steady stream without large bursts. This allows telephony engineers to minimize the spare capacity they have to set aside to handle usage bursts.

Another important property of telephone traffic is its high locality; most calls tend to be confined to geographical regions. This means that the capacity of long-distance networks (backbones) needed to provide adequate service is far less than the combined capacity of local exchanges. It is not entirely clear if the locality is a natural phenomenon (for example, long-distance calls for pizza delivery do not make a lot of sense) or an artifact of distance-sensitive pricing.

Internet traffic has an entirely different character. It was recently shown that Internet traffic does not obey the same rules as a Poisson stream; rather it has a fractal, or self-similar nature (at least in time scales most relevant to congestion control techniques). Self-similarity of the traffic pattern means that no matter at which time scale it is analyzed, it looks the same. In practical terms, it means that the probability of long bursts is not very different from the probability of short bursts; i.e. no matter how much spare capacity there is, as long as the capacity of the backbone is less than the combined peak capacities of customer connections, there will be periods of congestion. Unlike Poisson streams, the aggregation of self-similar streams does not produce better-behaving traffic.

The self-similarity of Internet traffic is most often explained as the result of communications not being truly independent. For example, transferring a group of files with FTP causes every new TCP session to be created as soon as previous one ended. The same behavior can be observed in e-mail delivery, or access to WWW documents.

Recent research suggests that not only packet traffic, but also the TCP session arrival pattern is self-similar. This means that if a switched virtual circuit network is to be substituted for a connectionless TCP/IP, the network will have to cope with periodical overloads of control units of its switches.

To make matters worse for network engineers, Internet traffic is non-local; i.e. Internet communications do not tend to be confined within geographical regions (although language barriers somehow increase locality of traffic on international scale). This means that practically all traffic generated by local Internet service providers is going through nation-wide providers' backbones.

References:

  • Vern Paxson, Growth Trends in Wide-Area TCP Connections, 1994.
  • Vern Paxson and Sally Floyd, Wide-Area Traffic: The Failure of Poisson Modeling, LBL-35238, 1994.

INDEX