The aim of this testcase was to measure the throughput of an SCTP association where the saturated sender and the reveiver were on the same machine. We measured the throughput for different user message sizes ranging from 10 bytes to 16500 bytes. The sender always sent a million messages. The Nagle algorithm is used.
Please note that the measurement for Linux is not complete because of the inability to handle small messages. The LKSCTP developers are investigating this problem.
The two curves for BSD show the impact of the CRC32C computation. BSD and Solaris with default settings do not compute the checksum. The results show that Linux can increase the performance if the checksum calculation for the loopback interface is also disabled.
It should also be noted that the Solaris implementation shows a throughput that is not as stable as the one BSD shows. This is more visible in the next figure which concentrates on user message sizes of 10 to 1500 bytes. Sun is investigating the problem of stability of the throughput.
