Figure 6 shows the throughput versus message length for Comms1 for the three packet sizes and the T9000 transputer. The maximum throughput reached for long message lengths is 8.3 Mbytes/s. This is less than the maximum C101 throughput of 9.2 Mbytes/s. The reason for this is a limitation in the C101 receive FIFO. Only in some very specific cases, the limitation allows achieving a throughput higher than 8.3 Mbytes/s. The peaks in the 4096 packet size graph show such a situation.
The minimal packet size required to achieve the maximum throughput in Comms1 is 512. Using packet size 256 the 8.3 Mbytes/s is nearly reached. The other packet sizes are too small to be handled efficiently enough to reach the maximum throughput.
The plots in the graphs show a sawtooth pattern. This pattern is related to the packet size: if the message length exactly fits within a number of data packets, the packet handling overhead is relatively low, and therefore the throughput will show an optimum. If, however, the message length is one byte longer, an extra data packet is needed, so the packet handling overhead will be relatively high, which causes a throughput dip.
The maximum Comms1 throughput achieved on two T9000 transputer depends on the network latency, caused by the C104 switches that interconnect the two transputers. Directly interconnecting the T9000 transputers will result in an increase of maximum throughput of about 25 %. Due to the use of a different acknowledgement scheme, interconnecting two DSNICs via a number of C104s does not affect the maximum throughput, it only affects the latency.
Figure 7 shows the throughput versus message length for Comms2. As expected, the maximum throughput reached for long messages is 16.6 Mbytes/s: twice the 8.3 Mbytes/s Comms1 throughput. This proves that the DSNIC can fully exploit the bidirectional bandwidth of DS links.
The minimal packet size required to achieve the maximum throughput in Comms2 is 512. Packet size 256 definitely does not achieve the maximum throughput. Considering the Comms1 situation, this is to be expected: since the throughput is twice as high, the packet size must roughly be doubled in order to keep the CPU load the same.