3.3 THE ADVANCED RESEARCH PROJECTS AGENCY NETWORK (ARPANET) COMPUTER RESOURCE SHARING WITH THE ARPA NETWORK (Lawrence G. Roberts, Director for Information Processing Techniques, Department of Defense, Advanced Research Projects Agency) Since its conception in 1968 the ARPA Network has progressed rapidly to become a major national computer resource sharing service interconnecting over 34 computer centers. Besides serving as the vehicle for the development and exploitation of multi-computer resource sharing, it has demonstrated the effectiveness and economy of an important new form of communications technology, packet-switched communications. Within the next five years, the extension of packet communication concepts into the areas of radio and satellite communication should revolutionize as well the effectiveness and economy of both local and international data communications. The current ARPANET was not designed for these extremes and is most efficient for interconnecting moderate to large computers or user complexes scattered at a fair density over a nationwide geographical area. Since the future will bring vastly improved techniques for both tying isolated small users to the network and interconnecting networks together worldwide, this presentation will concentrate on delineating the domain within which the ARPANET is used and for which its design makes it most economic. Historical development The first four ARPANET nodes were installed in late 1969 after a year and a half of study and development. It was in the initial period (1968) that most of the global design decisions were made; i.e., to use packet switching, interconnecting small Interface Message Processors (IMPs) with 50 KB lines, accepting messages of up to 8000 bits from capable host computers and breaking these messages into packets of 1000 bits or less for store and forward transmission. The initial analysis of network performance and cost, made in late 1968 just prior to contracting for the development of the IMPs, has proven to be very accurate and the same numbers are still quoted for parameters like the end-to-end packet delay (.1 sec). The second year (1969) was spent developing the IMPS and installing a four node test network. Since the initial network worked essentially as predicted and it was recognized that at least 15 nodes would be necessary to achieve adequate resource diversity and usefulness, the network was immediately expanded to include the 15 ARPA computer research centers most likely to contribute to the development of resource sharing software. The communication network tying these nodes together was complete by February 1971; however, the development of an agreed upon standard for host-to-host intercommunication and the related host software for all the computers was not complete until August 1971. Development of network usage Although very limited usage of the network existed throughout its early development, it was not until all fifteen nodes were compatibly interconnected in mid-1971 that true user activity could usefully begin. Shortly after this point, in September 1971, another phase of network growth was begun: the addition of user nodes without their own major host computer through the use of a new device, the Terminal Interface Processor (TIP-essentially a small IMP connected to its own minicomputer terminal processor). The TIP permitted a whole new class of network use, i.e., a group or center on the network could now look to the network for all its computation requirements instead of operating a local computer service. This strategy for obtaining the opimum mix and balance of computer capability from large, cost-effective computer centers without the many problems associated with running a local center has been so successful that about half the network nodes are now of this type. Also, during the past year several additional host computer service centers have been added to the network as the user requirements for quality computer service expanded. The entire expansion of the network beyond the original 15 research nodes has been in response to requirements (and full reimbursement) from non-network research projects, both in ARPA and other government agencies. Usage of the network has mainly been of three types: remote access to timeshared systems, subroutine-like use of large numerical machines, and file transfer. The remote access use is similar to the use of commercial time-sharing networks; users access the net through a TIP and use one of the many time-shared host computers on the net. For reliability they usually keep their files at several compatible host sites and use whichever host is available. The software and file transfer protocol has now developed to the point that if one is operating on one computer with files stored at another, it is both convenient and fast (seconds) to access the files. The second major usage of the net is to run large numerical processes on large, fast remote hosts while using a time-sharing computer for the interactive portions of the tasks (program editing, data input, output processing and display). For many applications, interactive manipulation of the data is essential but the central task of numerical computation can be run far faster and far cheaper on a large number cruncher host. The cost and time savings are often both more than a factor of ten. The third major network activity is file transfer. Files must be interchanged between host computers quite frequently for the pervious two applications but in addition there has been considerable use of the network for the bulk movement of large data and program files in preference to mailing tapes since the network is more reliable, less error-prone, usually cheaper, far faster and eliminates human handling. A six-month Air Force test of the network for the pure movement of data traffic showed that throughput rates of 20–30KB could be maintained. (Forthcoming changes to the routine technique should more than double this.) The test results also showed a lower monthly cost, while providing at least ten times the throughput and responsiveness of the alternatives available. So it has turned out that this type of network usage (pure data movement) is both economic and attractive even though the original design goals of the network were aimed at interactive computing. For the network, such data traffic is useful since it provides a large low-priority background load which expands the instantaneous capacity available and helps maintain efficient line utilization, thus reducing the cost to everyone. Network traffic Figure 1 shows the internode network traffic over the first fourteen months of user access. Usage has increased exponentially at the phenomenal rate of 26% per month throughout this period while network size has only increased linearly by one node per month. Inter-node traffic in October 1972 was 1.36 million packets per day which corresponds to 9% of the computer power in the network being used remotely. In addition to this, the network also handled .45 million packets per day of local traffic, which means that 12% of the total computing power of the network's 22 serving host computers was distributed via the network. On an annual basis the value of this computer time would be $2.1M or slightly more than network cost. However, at the current growth rate the network should be fully loaded by July 1973, in 9 months. Since the cost saving incurred by selecting the proper network computer for each problem is usually 100 to 300%, the network is already cost-effective. When it is fully loaded, the network costs only amount to 10-15% of the computer costs. In order to fully support computer resource sharing reliably and responsively on a nationwide basis, the communications network must be approximately the size and cost of the current ARPANET (34 nodes and $2M/yr). Based on recent measurements the network traffic generated by a fully loaded, moderate size, time-shared computer is 720,000 packets per day. A minimal network such as the current ARPAÑET has a basic capacity of 10M packets per day. Therefore, to fully utilize the basic capacity of the ARPANET would require 14 moderate size computers to be fully accessed through the network. Additional capacity can easily be added beyond this point with no great economy of scale. Figure 2 shows this effect for generalized national networks of ARPANET technology. By utilizing the above measurement of network traffic produced by a host and estimating its rental at $720,000 per year, total network traffic can be related to the total computer resource value accessed via the network (each computer dollar produces 365 packets of traffic). Thus the relative cost of network communication can be related to overall the annual computer value used. Since there are some fixed costs associated with adding each additional node, the cost-effectiveness depends partially on the number of nodes, but the main effect is produced by the dollar volume of usage. This means that at least 10-20 million dollars of computer time usage must be expected before a nationwide network becomes optimally cost-effective. Once the activity level is reached the main benefit from increased usage is improved reliability and increased peak throughput capability. |