: Network Throughput Test
In the Network Throughput test, one node (a producer) generates a message (a token) and then sends it to a second node (a consumer), which then returns the token to the producer. A node is either a producer or a consumer, never both, and is a member of, at most, one producer-consumer pair. The network load in this experiment was increased by increasing the token size over a range from four bytes to eight kbytes, and by increasing the number of producer-consumer pairs over a range from one to seven. Several runs consisting of several thousand exchanges were performed for each combination of token size and number of producer-consumer pairs. For each combination, the best observed throughput for the system was recorded. Throughput was measured as the number of tokens written multiplied by the token size and divided by the elapsed time. The results of this experiment are presented in Figure 3.
Not surprsingly, the local-wire experiments achieved the highest throughput, peaking at 6.3~MB/s f or an 8 KB token size. This is almost a factor of four improvement over the original bonded dual net Beowulf configuration. The remote-wire experiments succeed in bettering the original Beowulf performance, showing almost a factor of 2 improvement, but are unable to match the local-wire experiments. This is not unusual because the remote-wire experiments require that packets be duplicated through either software or switch routing mechanisms,introducing a latency penalty caused by contention for network resources and packet routing overhead." Unexpectedly, the switched channel bonded configurations did not exceed the performance of the equivalent switched non-channel bonded schemes, actually demonstrating marked degradation in throughput. These results are further discussed in Section 5.
The Switched Mesh Local-Net and the Routed Mesh Local-Net experiments are operationally identical and exhibit the same behavior. Because in both cases all transactions are between nodes on the same local wire, no messages must be passed through either a switch or a routing node. In other words, all transactions are of the type between nodes B8 and B9 in Figure 1 and nodes BA and BE in Figure 2. Therefore all transactions avoid the additional latency incurred by packet duplication through a switch or routing node in reaching a remote node.
