414 Part 2 ½ Regions of Computer Space
Section 5½ Networks
 

Table 2 Transit Times and Message Rates
 

	Minimum			Maximum
Single word message
Transit time  Round-trip  Max. message rate/link			5 msec 10 msec 100/sec		50 msec 100 msec 10/sec
Single full packet message
Transit time  Round-trip  Max. message rate/link			45 msec 50 msec 20/sec		140 msec 190 msec 5/sec
8-packet message
Transit time  Round-trip  Max. message rate/link		265 msec 195 msec 5/sec			360 msec 320 msec 3/sec

rate per link, assuming the negligible queueing delay of a lightly loaded net. In this table, "minimum" delay represents a short hop between two nearby IMPs, and "maximum" delay represents a cross-country path involving five IMPs. In all cases the delays are well within the desired half-second goal.

In a lightly-loaded network with a mixture of nearby and distant destinations, an example of heavy Host traffic into its IMP might be that of 20 links carrying ten single-word messages per second and four more links, each carrying one eight-packet message per second.

Computational Load

In general, a line fully loaded with short packets will require more computation than a line with all long packets; therefore the IMP can handle more lines in the latter case. In Fig. 11, we show a curve of the computational utilization of the IMP as a function of message length for fully-loaded communication lines. For example, a 50-kilobit line fully loaded in both directions with one-word messages requires slightly over 13 percent of the available IMP

time. Since a line will typically carry a variety of different length packets, and each line will be less than fully loaded, the computational load per line will actually be much less.

Throughput is defined to be the maximum number of Host data bits that may traverse an IMP each second. The actual number of bits entering the IMP per second is somewhat larger than the throughput because of such overhead as headers, RFNMs, and acknowledgments. The number of bits on the lines is still larger because of additional line overhead such as framing and error control characters. (Each packet on the phone line contains seventeen characters of overhead, nine of which are removed before the packet enters an IMP.)

The computational limit on the IMP throughput is approximately 700,000 bits per second. Figure 12 shows maximum throughput as a function of message length. The difference between the throughput curve and the line traffic curve represents overhead.
 
 

Discussion

In this section we state some of our conclusions about the design and implementation of the ARPA Network and comment on possible future directions.

We are convinced that use of an IMP-like device is a more sensible way to design networks than is use of direct Host-to-Host connection. First, for the subnet to serve a store-and-forward role, its functions must be independent of Host computers, which may often be down for extended periods. Second, the IMP program is very complex and is highly tailored to the I/O structure of the DDP-516; building such complex functions into special I/O units