Description
Datakit is a highly modular virtual-circuit switch that is designed to act as a component of a universal data transport system. The network architecture is structured to provide efficient data transmission in both wide-area and local-area environments. AT&T Bell Laboratories has established a nation-wide Datakit research network called XUNET, to study Datakit performance in the wide-area context. Our study utilizes this network to characterize the performance of remote executions on UNIX hosts.
We designed a number of user-level and network-level tests to identify and characterize the host and network delay factors that contribute to remote command execution delay, as seen by the user. This delay is composed of call-processing delays within the network nodes during call setup, processing time at the destination host, and call-processing delays within the network nodes during call teardown. We found that mean call-processing delay during both call setup and call teardown is roughly linear with the node count of the virtual-circuit path that connects the source and destination hosts. The mean call setup delay is 0.30 seconds per node, while the mean call teardown delay is 0.15 seconds per node.
We determined that the delay encountered at the remote host is composed of (1) the processing time of a UNIX server process, which handles Datakit call requests on XUNET hosts, and (2) the execution time of one or more processes created by the server to execute the command(s) specified for remote execution. In general, these host delays are more difficult to quantify than the network delays because host load fluctuations during testing tend to introduce unpredictable distortions in the results.