Datakit, a virtual-circuit switch developed at AT&T Bell Laboratories, is intended to provide efficient data communication over both local-area and wide-area configurations. In the wide-area case, its switching engine and the queuing disciplines it employs at the trunk interfaces try to balance the opposite requirements of low-delay interactive traffic and high-throughput data-transfer traffic. In order to experiment with distributed applications over Datakit technology, AT&T Bell Laboratories has installed an Experimental University Network (XUNET) that connects computers of the Berkeley Computer Science Division to machines at other universities and to machines at Bell Laboratories locations in New Jersey through 1.5 Mb/s T1 lines and Datakit nodes.
We measured the delay and throughput performance of XUNET. The window size and the small memory available in each host interface were the primary factors affecting the throughput. The queueing delays at each host affected the delays for small messages. We isolated and characterized potential bottlenecks in the network, and other sources of delay, to enable the network's designers to improve its performance. We provided a mathematical model for the delay and throughput characteristics of long-distance Datakit communication. The parameters of the model included protocol window size, source and destination distance, and line and interface speeds.