Untitled topic

I've been working on congestion control for a while, and one thing that has always bothered me is how traditional loss-based algorithms such as Reno and CUBIC treat packet loss as a congestion signal. There are also many delay-based approaches using RTT, but I wanted to try a different direction: combine loss and RTT into a single probabilistic signal that estimates how congested the network currently is, rather than treating congestion as a binary event.

To experiment with this idea, I built SCP, a reliable transport protocol over UDP, and designed a congestion-control algorithm called SCP-PROB.

SCP-PROB uses deviations in RTT and packet loss to compute a nonlinear congestion probability, then continuously adjusts its sending behavior based on that estimate.

In my experiments, it performs better than Linux CUBIC and other loss-based CCAs in high-random-loss environments, while multiple SCP-PROB flows also show good fairness with each other.

The implementation and experiments are still evolving, and there are several open problems—especially fairness against heterogeneous CCAs and how to attribute congestion to individual flows.

Here's the paper. Criticism is very welcome.