We check the similarity of Composable UDT by comparing CTCP and Linux TCP. Over a 1Gb/s link, we start 1 to 64 flows of CTCP and Linux TCP respectively. We record the throughput, fairness index, and stability index, which covers the majority of a transport protocol's performance properties.
This table lists these values for both CTCP and Linux TCP. We can see that both implementations have similar performance. However, CTCP does have a CPU usage overhead, as we can see that at 64 parallel flows, CTCP has smaller throughput than Linux TCP.
The overhead comes from additional memory copies, context switches, and user thread synchronizations. This is especially significant at the sender side.
This is one shortcoming of Composable UDT, but we argue that in many situations, this is not a problem. For example, this overhead only exists when comparing it to kernel implementations; other user space implementations have the same overhead too.