Loopback resultsOur aim was to compare the different SCTP kernel implementations in terms of their throughput. We chose the operation systems
The tests were performed on six identical machines with Intel Pentium IV 2.4 GHz processors. As network adapters we used Intel Pro/1000MT Gigabit server adapters.
We not only wanted to evaluate the implementations on their own but also in interoperation with the other systems. So every OS was sender and receiver of the data of every other system. This resulted in multiple different tests.
We tested the throughput at the user layer of all operating systems where the sender and receiver run at the same system for different user message sizes. The results show that the throughput depends on the user message length and that Solaris and FreeBSD perform substantially better, because they do not compute the CRC32C checksum.
We tested the throughput at the user layer where the sender runs on one host and the receiver runs on another host for different message sizes. All systems were connected directly via a Gigabit Switch with an MTU of 1500 bytes. We tried all combinations of operating systems but we figured out that currently Linux is not able to act well as a receiver. Therefore we can not provide these measurements.
