: Disk/Net Balance Test
In the Disk/Net Balance test, a producer node copies a file from its local disk (the Linux buffer cache is flushed before each copy) to a consumer node. The consumer can be either the producer node itself (an intranode copy) or a separate node in the system (an internode copy). Each node is a member of a maximum of one producer-consumer pair. To evaluate the degree to which the system disk and network subsystems achieve balanced performance over a range of conditions, measurements were made with seven producer consumer pairs for all possible combinations of intranode and internode file copies. This series of measurements was performed with files sizes ranging from one to 16 MBytes. Again, system throughput was measured as the number of files copied times the files size, divided by the time needed for all copies to complete. Because of space considerations, we only present the results for the 2 MB file copies in Figure 4 to elucidate the performance differences of the studied network configurations.
The disk I/O experiments mirror the results of the network throughput tests, demonstrating the overall poorer performance of the channel bonded network configurations and the superiority of local-wire transactions over remote-wire ones. With respect to interprocessor copies, a peak sustained throughput of about 6 MB/s was achieved by both the software routed and switched local-wire tests, while the channel bonded configurations were able to peak at 7.2 MB/s for one remote file copy, before degrading. The one remote file copy data shows more clearly than the network throughput tests channel bonding's ability to better handle small workloads. These data are discussed in more detail in the following section.
