Chapter 40
Computer network examples
We are just entering the era in which general-purpose networks of computers make technical and economic sense. The requisite hardware and software development of operating systems and multiprogramming capability is still maturing. Thus, unlike the other PMS structures discussed in this book, there is no supply of operational systems with published descriptions upon which we can draw. Consequently, we have assembled several brief examples of networks to provide at least some illustrations of what is sure to be an important aspect of computer systems in the near future. The more interesting of these examples are still in the planning stages; those that exist currently are still highly specialized.
Spatially distributed intercommunicating networks of digital devices have existed for a long time. But many of the ones that come most easily to mind are not computer networks. For example, the various airline reservation systems like American Airline's SABRE [Plugge and Perry, 1961] have spatially distributed terminals (T's) with a single Pc, possibly mediated by Pio's or Cio's. When there are several Pc's, they are functionally integrated so as to provide the total capacity and reliability needed. Some military networks, such as the SAGE Air Defense System [Everett et al., 1957] have multiple computers (SAGE actually has a very large number). But they transmit to each other highly specialized data streams (for example, aircraft positional information for control). The National Physics Laboratory of England has made a very comprehensive proposal for a general-purpose network [Davies et al., 1967], although we do not include it as a chapter. Again, it is just in the proposal stage. The Lawrence Radiation Laboratory (at Livermore) is no doubt the earliest and most impressive network.
In terms of our PMS descriptions, a computer network (N) requires at least two C's not connected through primary memory. Thus each C has a Pc and an Mp of its own and has to communicate with other C's through messages. Duplex computers are thus defined as networks, provided they do not share Mp. For networks, links (L's) are usually shown explicitly. In spatially distributed systems, both the time delays and the flow rates of the links are significant. The latter is so partly because the networks must make use of the telephone communication system, which exists independently of the networks, thus having parameters that do not correspond with any of the internal parameters of the individual computers. There may also be limitations of reliability, cost, accessing characteristics, and the size of the information unit that derive wholly from the links. For instance, many computer networks would like to buy their transmissions from the telephone system for very short intervals (milliseconds), at very high data rates, and with short switching time (milliseconds), i.e., bursts. Switching time and pricing policies within the telephone system conspire to make this a difficult thing to do. Thus, with networks, links become important independent components.
One classification of networks (N's) is by fixed or variable interconnection structure. Fixed structure may mean that the links are fixed permanently over the life of the network. However, fixed structure may mean only that connections once made must be held for long periods of time relative to the message flows. An example is the telephone switching system mentioned above, which looks like a variable switching structure at the level of human conversations, but like a fixed switching structure at the level of computer conversations. Figures 1a and 1c show variable-structure systems; Fig. 1b shows a fixed-structure system. In the former, any C can talk directly to any other C. In the latter, each C talks directly to only a few C's; thus, to communicate with the other C's, it must transmit through them as links; that is, it must use another C as an L.
A second classification of N's is by the nature of the delays suffered by the messages as they travel from an initiating C to a target C. Communication can be direct, in which case the only delays are those through the switches (S) and links (L) between the two C's (Figs. 1a and 1b). Alternatively, communication can involve storing messages at intermediate nodes (called store-and-forward communication), thus introducing additional memory delays into the communication but decreasing the demands for coordination between the two C's. Although store-and-forward systems can be built with the intermediate nodes being K's with buffer memories, in the present context the natural form for such a system uses the other C's in the system as the intermediate nodes, as in Fig. 1c.
Several kinds of reasons can justify the existence of a particular network. The following list is adapted from Roberts [1967]:
Load sharing. A problem (program and data) initiated at one C that is temporarily overloaded is sent to another for processing. The cost of transshipment must clearly be less than the costs of
504