In this section, methods used to evaluate the relative communications capacity of five alternative network configurations are described in conjunction with the resulting empirical data. A synthetic program, the Network Throughput test, and a parallel file copy application, the Disk/Net Balance test, were run on each network configuration to make the comparison. Both experiments used TCP/IP as the networking protocol, and results were collected on a node not otherwise involved in the experiment.
With all of the mesh configurations evaluated here, the relationship between a pair of nodes falls into one of two categories. When the members of a pair of nodes share one of the two networks' segments to which they are connected, communication between them is referred to as a local-wire transaction. Alternatively, when the members of a pair of nodes share no common network segments, communication between them is referred to as a remote-wire transaction. For those network configurations where the local-wire/remote-wire distinction was meaningful (all segmented topologies), a complete set of measurements was collected for local-wire only and remote-wire only transactions. In all cases, the experiments were designed to avoid contention between transactions to the extent possible.
Three principal features describe the experiments that were conducted in evaulating the different network configurations:
These situations are encapsulated by Figures 1 and 2. The figures themselves are differentiated by routing system, Figure 1 representing software routing and Figure 2 switch based routing. The two packet paths from B0 to B5 in Figure 2 show channel-bonding as well as exemplifying remote-wire transactions. In Figure 1, the B8 to B9 packet path demonstrates a local-wire transaction.